GloGen BP Estimation

Créé le : 14 décembre 2024

Répondu en utilisant O1 par Chat01

Question

详细介绍这篇文章,完整详细Contents lists available at ScienceDirect
Computers in Biology and Medicine
journal homepage: www.elsevier.com/locate/compbiomed
GloGen: PPG prompts for few-shot transfer learning in blood pressure
estimation
Taero Kim a
, Hyeonjeong Lee c
, Minseong Kim c
, Kwang-Yong Kim c
, Kyu Hyung Kim c
,
Kyungwoo Song a,b,∗
a Department of Statistics and Data Science, Yonsei University, 50, Yonsei-ro, Seodaemun-gu, 03722, Seoul, Republic of Korea
b Department of Applied Statistics, Yonsei University, 50, Yonsei-ro, Seodaemun-gu, 03722, Seoul, Republic of Korea
c Daegu-Gyeongbuk Research Division, Electronics and Telecommunications Research Institute (ETRI), 1, Techno sunhwan-ro 10-gil, Yuga-eup,
Dalseong-gun, 42994, Daegu, Republic of Korea
A R T I C L E I N F O
MSC:
41A05
41A10
65D05
65D17
Keywords:
Photoplethysmogram signal
Prompt learning
Few-shot learning
Transfer learning
Blood pressure estimation
A B S T R A C T
With the rapid advancements in machine learning, its applications in the medical field have garnered increasing
interest, particularly in non-invasive health monitoring methods. Blood pressure (BP) estimation using
Photoplethysmogram (PPG) signals presents a promising opportunity for real-time, continuous monitoring.
However, existing models often struggle with generalization, especially for high-risk groups like hypotension
and hypertension, where precise predictions are crucial. In this study, we propose Global Prompt and Prompt
Generator (GloGen), a robust few-shot transfer learning framework designed to improve BP estimation using
PPG signals. GloGen employs a dual-prompt learning approach, combining Global Prompt (GP) for capturing
shared features across signals and an Instance-wise Prompt (IP) for generating personalized prompts for each
signal. To enhance model robustness, we also introduce Variance Penalty (VP) that ensures diversity among
the generated prompts. Experimental results on benchmark datasets demonstrate that GloGen significantly
outperforms conventional methods, both in terms of accuracy and robustness, particularly in underrepresented
BP groups, even in scenarios with limited training data. GloGen thus stands out as an efficient solution for
real-time, non-invasive BP estimation, with great potential for use in healthcare settings where data is scarce
and diverse populations need to be accurately monitored.

  1. Introduction
    BP monitoring is a critical component in managing cardiovascular
    health, providing essential insights into heart function. Traditional BP
    measurement methods are effective but require cumbersome equip￾ment. It makes continuous BP monitoring an inconvenience for patients
    who require it. Considering that BP is a crucial metric for assessing
    individual health status, the implementation of real-time BP monitoring
    via wearable devices holds the potential to yield significant economic
    and healthcare benefits. As a result, BP estimation has spurred interest
    in developing more convenient and non-invasive techniques, leading to
    the exploration of utilizing Photoplethysmogram (PPG) signals for BP
    estimation.
    PPG signal is a biosignal that indicates the changes of degree of light
    absorption based on the blood volume change or the velocity of blood
    flow. Numerous studies have extensively investigated the correlation
    between PPG and Arterial Blood Pressure (ABP) signals that contain
    ∗ Corresponding author at: Department of Applied Statistics, Yonsei University, 50, Yonsei-ro, Seodaemun-gu, 03722, Seoul, Republic of Korea.
    E-mail addresses: [email protected] (T. Kim), [email protected] (H. Lee), [email protected] (M. Kim), [email protected] (K.-Y. Kim),
    [email protected] (K.H. Kim), [email protected] (K. Song).
    actual BP information [1,2]. Unlike ABP signals, PPG signals are mea￾sured non-invasively, allowing for blood pressure monitoring without
    temporal or spatial constraints, which holds significant potential for
    widespread impact. In pursuit of more precise BP estimation, there is
    a burgeoning body of research focused on leveraging deep learning
    techniques to either estimate systolic and diastolic blood pressure (SBP
    and DBP, respectively) or to generate ABP directly [3–7]. One of the
    challenges in BP estimation is that the model’s predictive performance
    varies significantly across subgroups, such as disease, ethnicity, and
    gender [8,9]. In particular, for high-risk groups like patients with
    hypotension or hypertension, where precise BP measurement is more
    critically required, the performance tends to be somewhat less sat￾isfactory compared to individuals with normal blood pressure [10–
    12].
    As vast amounts of data accumulate, the field of machine learning
    has demonstrated notable success with transfer learning. It involves
    https://doi.org/10.1016/j.compbiomed.2024.109216
    Received 31 March 2024; Received in revised form 21 September 2024; Accepted 25 September 2024
    Computers in Biology and Medicine 183 (2024) 109216
    Available online 8 October 2024
    0010-4825/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
    T. Kim et al.
    initially learning the general features of substantial data and then
    applying this knowledge to smaller datasets specific to a target task.
    Similarly, in the realms of medicine and healthcare, advancements in
    database systems like Electronic Health Records have enabled access
    to extensive medical datasets, such as MIMIC. Recently, these datasets
    have been used to evaluate diverse BP estimation algorithms using
    PPG, along with the construction of various benchmark datasets [13].
    However, research on the efficient transfer learning methods for BP
    estimation using these resources is not being actively conducted. Partic￾ularly, effectively applying models to target tasks requires specific data
    related to those tasks, but acquiring sufficient data for the target tasks
    is challenging due to the nature of medical data. Therefore, exploring
    desirable transfer learning methods that achieve better generalization
    ability for BP estimation and robustness across BP groups with a limited
    number of data is essential.
    In this study, we propose Global Prompt and Prompt Generator
    (GloGen), a novel few-shot transfer learning algorithm for BP esti￾mation using PPG signals. GloGen consists of Global Prompt (GP)
    composed of trainable parameters itself, and Instance-wise Prompt (IP)
    generated by Prompt Generator. Uniquely, GloGen combines two types
    of prompts into a PPG signal. This dual-prompt strategy allows our
    model to capture both common features across PPG signals and distinct
    features of individual PPG signals. Consequently, IP applies different
    augmentation policies for each instance, ensuring robust estimation
    by the model even for underrepresented BP groups in the pre-trained
    models. Moreover, we introduce Variance Penalty (VP) to encourage
    diversity among IPs. VP prevents the model from bias towards certain
    BP groups or generating similar IPs. By promoting the diverse prompts,
    we facilitate more personalized PPG signals, crucial for robust BP
    estimation especially in hypotensive or hypertensive BP groups.
    To ensure a fair evaluation of our method, we conduct experiments
    with reconstructing recently proposed benchmark datasets [13]. To
    evaluate the model robustness across different BP groups, we divide
    each instance according to specific criteria for assigning BP group and
    report both the performance for each group and the group average per￾formance, which is the average of these group-specific performances.
    It allows us to quantitatively assess whether our model performs BP
    estimation robustly across BP groups.
  2. Related works
    2.1. PPG-based blood pressure estimation
    PPG signals reflect the variations in the amount of light reflected
    due to differences in blood volume. One reason BP estimation based
    on PPG is actively studied is that PPG is a biometric signal that can be
    obtained non-invasively while showing a correlation with ABP signals,
    which are invasive [1,2]. Traditional ML methods like Random For￾est [14] extract features known to have a strong correlation with blood
    volume from PPG signals for BP estimation [15–17]. Additionally, some
    studies not only utilize PPG signals but also employ additional biosig￾nals containing other information on BP, such as Electrocardiogram
    (ECG) signals. [18,19]. There is also active research on generating ABP
    directly from PPG signals for continuous BP estimation [8,9,20,21].
    However, this study aims to directly estimate SBP and DBP using only
    PPG signals [4,22,23]. [13] evaluate various models for BP estimation
    through PPG signals and introduce various benchmark datasets for fair
    evaluation. Unlike previous studies, our research presents a transfer
    learning strategy that effectively utilizes the knowledge of a pre-trained
    model for target tasks. We also assume a few-shot setting where only a
    limited number of data are available for the target task. Moreover, in
    addition to the overall BP estimation performance, we introduce evalu￾ation metrics that reflect the differences in BP estimation performance
    across BP groups to assess the robustness of the model.
    2.2. Prompt learning
    Prompt learning, first introduced in NLP, serves as an alternative
    to the costly fine-tuning of pre-trained language models, utilizing a
    small number of additional parameters for model training [24,25].
    Prompt learning trains the optimal prompt to enable the model to adapt
    to the target task easily, demonstrating superior generalization per￾formance compared to fine-tuning [26,27]. In Computer Vision (CV),
    prompt learning has shown promising results in various recognition
    tasks by introducing trainable parameters and adding them to the
    input images [28,29]. By learning an optimal augmentation policy,
    prompt learning potentially restructures the input images to enable the
    model to achieve better generalization ability on the target task [30].
    Similarly, research in Vision–Language Models [31,32] and time-series
    forecasting [33] is gaining momentum. This study demonstrates that
    prompt learning is an effective transfer learning strategy for PPG signals
    as well. GloGen generates PPG prompts in a manner that ensures ro￾bustness to BP groups as well as overall model predictive performance,
    even with a limited amount of data for the target task. That is, PPG
    prompts reformulate the PPG signals to make models more robust. They
    can be integrated with input data in various ways, and we opt for a
    method directly adding the PPG prompt to the input data, ensuring it
    functions well regardless of the model structure [29].
    2.3. Comparison with existing research
    Existing research for BP estimation using PPG signal has generally
    progressed by improving model architectures [4,9,23]. However, our
    study operates differently by generating new prompts using only the
    input and output without making direct changes to the model itself.
    This approach implies that it is applicable to any model, as long as
    the model’s embeddings can be obtained. Additionally, some studies
    require demographic information [8,9], but GloGen only needs SBP and
    DBP from the training dataset, and the input PPG signal, as BP groups
    are determined by SBP and DBP. During testing, GloGen does not use
    group information and relies on the model’s predictions. Furthermore,
    our approach using prompt learning assumes a few-shot setting, unlike
    previous PPG-related transfer learning studies [8,9]. Specifically, we
    perform transfer learning with only 5 or 10 samples per blood pressure
    group, which contrasts with previous research that requires the entire
    training set of the target dataset for transfer learning. We also observe
    and improve performance differences across blood pressure groups de￾fined by small intervals. Many studies on BP estimation using PPG focus
    solely on average BP prediction performance [13,34]. However, we
    concentrate on enhancing model robustness as well, achieving this by
    incorporating minimal additional parameters through prompt learning.
  3. Methodology
    3.1. Global prompt and prompt generator
    We introduce two types of prompts: Global Prompt (GP) and
    Instance-wise Prompt (IP). GP is a set of trainable parameters itself and
    is equally added to all input PPG signals. On the other hand, IP reflects
    the individual characteristics of each input PPG signal. Therefore,
    each instance generates a different IP, for which we construct Prompt
    Generator. When estimating BP based on PPG signals, the common
    characteristics of all PPG signals are important for overall generaliza￾tion ability. In our methodology, GP is considered as a prompt that
    learns such shared features. However, for data belonging to a certain
    BP group, it is crucial to learn the specific characteristics of each data.
    IP addresses this need, enabling our model to robustly handle underrep￾resented BP groups such as hypotension or hypertension groups in the
    pre-trained model. Ultimately, to ensure robust estimations for different
    BP groups as well as overall predictive ability, we utilize both GP
    Computers in Biology and Medicine 183 (2024) 109216
    2
    T. Kim et al.
    Fig. 1. Overview of GloGen model. 𝑥
    𝐺𝑃 and 𝑥
    𝐼𝑃 denote Global Prompt (GP) and Instance-wise Prompt (IP), respectively. Encoder ℎ uses PPG signals to get embeddings 𝑍𝑒𝑚𝑏𝑒𝑑 for
    generating IP by Prompt Generator 𝑔(𝜙). PPG signal, combined with 𝑥
    𝐺𝑃 and 𝑥
    𝐼𝑃 , passes through the pre-trained model 𝑓 = 𝑓𝜃
    𝑝𝑒𝑛
    ◦𝑓𝜑
    𝑟𝑒𝑔 for BP estimation. 𝑉 𝑃 and 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 represent
    the variance of IPs across BP groups and the trigger vector, respectively.
    and IP. We name this dual-prompt strategy Global Prompt and Prompt
    Generator (GloGen). Fig. 1 shows the overview of GloGen.
    In GloGen, Prompt Generator 𝑔𝜙
    corresponds to the decoder within
    the encoder–decoder architecture. To generate optimal prompts, Glo￾Gen offers two configuration options. The first pertains to the configu￾ration of the encoder ℎ. By default, as an encoder in GloGen, we adopt
    a feature extractor 𝑓𝜃
    𝑝𝑒𝑛 that corresponds to all but the last layer of the
    pre-trained model 𝑓 = 𝑓𝜃
    𝑝𝑒𝑛
    ◦𝑓𝜑
    𝑟𝑒𝑔 , i.e., ℎ = 𝑓𝜃
    𝑝𝑒𝑛, where 𝑓𝜑
    𝑟𝑒𝑔 stands for
    the linear layer for regression. This approach aims to maximize the use
    of the pre-trained model’s knowledge for generating prompts. However,
    if 𝑓 is undertrained or strongly biased, it may lead to suboptimal
    embeddings 𝑍𝑒𝑚𝑏𝑒𝑑 for prompt generation. For these reasons, Principal
    Component Analysis (PCA) encoder 𝑓
    𝑝𝑐𝑎 can be a viable option, that
    is, ℎ = 𝑓
    𝑝𝑐𝑎. PCA encoder constructs a fixed projection matrix for the
    PCA space using few-shot data from the training set. This ensures that
    during the inference stage, the test set is projected onto this PCA space
    to obtain 𝑍𝑒𝑚𝑏𝑒𝑑 , thereby also ensuring faster inference times.
    The second option pertains to the construction of embeddings. The
    embedding 𝑍𝑒𝑚𝑏𝑒𝑑 , derived from the pre-trained model, extracts fea￾tures of the target task data by solely considering the characteristics of
    the data utilized during pre-training. Consequently, it fails to generate
    appropriate embeddings for information unique to the target task.
    Inspired by [29], we introduce a randomly initialized trainable vector,
    referred to as the trigger vector 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟. Unlike 𝑍𝑒𝑚𝑏𝑒𝑑 , 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 is not
    instance-specific. For a target task, a single 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 is concatenated
    uniformly across all 𝑍𝑒𝑚𝑏𝑒𝑑 and then input into Prompt Generator.
    𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 is updated during the training stage and models the input
    vector for Prompt Generator that represents the embedding for the
    additional information to generate the appropriate prompt adopting
    models to the target task. Note that the role of GP and 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 differ.
    While GP aims to learn the shared characteristics across all the training
    data, 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟, in conjunction with instance-specific information 𝑍𝑒𝑚𝑏𝑒𝑑 ,
    strives to facilitate the generation of more optimal prompts.
    3.2. Construction of GloGen prompts
    Let 𝑖 denote the data index. Both GP, 𝑥
    𝐺𝑃 , and IP, 𝑥
    𝐼𝑃
    𝑖
    , share
    the same dimension as the input PPG signal, 𝑥𝑖
    . 𝑥
    𝐺𝑃 is a learnable
    parameter added uniformly to all 𝑥𝑖
    . Through this process, 𝑥
    𝐺𝑃 is
    expected to capture common features within the training data. Unlike
    𝑥
    𝐺𝑃
    , 𝑥
    𝐼𝑃
    𝑖
    is generated by Prompt Generator 𝑔𝜙
    :
    𝑥
    𝐼𝑃
    𝑖
    = 𝑔𝜙◦ℎ(𝑥𝑖
    ) = 𝑔𝜙
    (𝑧𝑖
    ), (1)
    where 𝑧𝑖 ∈ 𝑍𝑒𝑚𝑏𝑒𝑑 in the default setting and 𝑧𝑖 ∈ [𝑍𝑒𝑚𝑏𝑒𝑑 , 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟]
    when using trigger vector. 𝑔𝜙
    consists of a set of deconvolution layers
    (also known as Transposed Convolution layers), batch normalization,
    and ReLU activation function. IP, in contrast to GP, learns the unique
    characteristics that are associated with individual data or groups of
    data. Therefore, IP regulates each instance through corresponding per￾sonalized prompt, imposing additional constraints on challenging BP
    groups that are underrepresented in the model. We construct GloGen
    prompts by using GP and IP. Finally, the input adding GloGen prompt
    is:
    𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖
    = 𝑥𝑖 + 𝜆 ⋅ 𝑥
    𝐺𝑃 + 𝛾 ⋅ 𝑥
    𝐼𝑃
    𝑖
    ,
    where 𝜆 and 𝛾 stand for the coefficients controlling the effects of GP
    and IP, respectively.
    𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖
    serves as the input to the pre-trained model, which then
    predicts SBP and DBP for BP estimation. To train GP and Prompt
    Generator, we use Mean Squared Error (MSE) as a loss function:
    𝐿𝑀𝑆𝐸 =
    1
    𝑁
    𝑁∑
    𝑖
    (𝑓(𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖
    ) − 𝑦𝑖
    )
    2
    ,
    where 𝑁 is the number of training data and 𝑦𝑖
    is ground-truth SBP and
    DBP corresponding to 𝑥𝑖
    . Note that we always freeze 𝑓𝜃
    𝑝𝑒𝑛. Concretely,
    in the default setting, we only update the 𝑔𝜙
    , GP, and 𝑓𝜑
    𝑟𝑒𝑔 for recon￾figuration on the embeddings. If we use 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 or 𝑓
    𝑝𝑐𝑎, we also update
    𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 or use training data for the target task to get the projection
    matrix, respectively.
    3.3. Normalization
    As 𝑥
    𝐺𝑃 and 𝑥
    𝐼𝑃
    𝑖
    are trainable, they can take any value. If the value
    of the prompt is much larger than the actual PPG input, it overshadows
    the information of the original PPG input, leading to training instability
    or predictions. Therefore, we introduce the normalization using the
    statistics of the training dataset.
    Let 𝑥𝑖𝑗 denote 𝑗th component of 𝑥𝑖
    , 𝑥
    𝑃
    𝑖
    denote prompt vector, and
    𝑋𝑡𝑟 denote a set of input PPG signals in the training set. 𝑥
    𝑃
    can be any
    𝑥
    𝐺𝑃
    , 𝑥
    𝐼𝑃
    𝑖
    or 𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖
    . When max(𝑋𝑡𝑟) = max𝑖,𝑗 𝑥𝑖𝑗 and min(𝑋𝑡𝑟) = min𝑖,𝑗 𝑥𝑖𝑗
    denote the maximum and minimum value in 𝑋𝑡𝑟, respectively, the
    normalized 𝑥
    𝑃
    𝑖
    is:
    ̂𝑥
    𝑃
    𝑖
    =
    max(𝑋𝑡𝑟) − min(𝑋𝑡𝑟)
    max(𝑥
    𝑃
    𝑖
    ) − min(𝑥
    𝑃
    𝑖
    )
    (𝑥
    𝑃
    𝑖
    − min(𝑥
    𝑃
    𝑖
    )) + min(𝑋𝑡𝑟),
    where max(𝑥
    𝑃
    𝑖
    ) = max𝑗 𝑥
    𝑃
    𝑖𝑗 and min(𝑥
    𝑃
    𝑖
    ) = min𝑗 𝑥
    𝑃
    𝑖𝑗. By normalizing 𝑥
    𝑃
    𝑖
    and adding ̂𝑥
    𝑃
    𝑖
    to the original PPG, we aim to stabilize the training
    process and ensure more consistent predictions.
    Inspired by [29], we also consider clipping the GloGen prompt. This
    approach, akin to normalization, is known to enhance the stability of
    prompt learning. Clipping ensures that the GloGen prompt values are
    constrained within the range defined by min(𝑋𝑡𝑟) and max(𝑋𝑡𝑟):
    ̂𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 =





    min(𝑋𝑡𝑟) if 𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 < min(𝑋𝑡𝑟)
    𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 if min(𝑋𝑡𝑟) ≤ 𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 ≤ max(𝑋𝑡𝑟)
    max(𝑋𝑡𝑟) if max(𝑋𝑡𝑟) < 𝑥𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 .
    Computers in Biology and Medicine 183 (2024) 109216
    3
    T. Kim et al.
    Table 1
    Performance Evaluation of GloGen. From (a) to (f), each represents the performance according to the pair of the data used for pre-training and the target task data. We report
    average MAE and group average MAE for both SBP and DBP, respectively. The total average MAE (Avg) and total group average MAE (GA) refer to the sum of the respective
    metrics for SBP and DBP. We conduct all experiments for both 5-shot and 10-shot settings. In all experiments, we select the model based on the best validation total average MAE.
    2
    refers to the coefficient of determination. All units are mmHg.
    (a) Pre-trained: BCG - Target: Sensors
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 46.35 35.69 10.66 37.25 28.77 8.49 -4.11
    LP 38.94 28.43 10.51 34.79 25.60 9.19 -2.77
    FT 33.32 24.41 8.92 31.56 23.52 8.03 -1.55
    GloGen 32.47 23.70 8.77 30.20 22.84 7.36 -1.42
    10
    Scratch 35.81 26.11 9.70 32.75 24.12 8.62 -2.02
    LP 33.06 24.07 8.99 31.26 23.46 7.80 -1.54
    FT 32.72 24.25 8.46 30.78 23.21 7.57 -1.39
    GloGen 31.71 22.80 8.91 29.76 22.35 7.40 -1.35
    (b) Pre-trained: BCG - Target: UCI
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 32.65 22.08 10.57 33.68 23.69 9.99 -1.74
    LP 28.89 19.18 9.71 32.09 23.16 8.94 -1.02
    FT 28.19 19.15 9.04 32.10 22.73 9.37 -0.77
    GloGen 27.75 19.09 8.67 30.23 22.29 7.94 -0.72
    10
    Scratch 30.03 20.78 9.25 31.21 22.84 8.38 -1.16
    LP 26.09 17.37 8.72 31.37 22.72 8.64 -0.45
    FT 27.58 18.81 8.77 31.84 22.57 9.27 -0.64
    GloGen 25.74 17.63 8.11 30.55 22.39 8.16 -0.30
    (c) Pre-trained: Sensors - Target: BCG
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 25.35 14.23 11.13 34.26 22.45 11.82 -1.75
    LP 22.06 12.58 9.48 31.62 20.40 11.22 -0.94
    FT 30.28 18.95 11.34 42.02 28.32 13.70 -3.00
    GloGen 19.97 11.87 8.10 27.92 17.98 9.94 -0.30
    10
    Scratch 20.09 12.56 7.53 32.26 22.56 9.70 -0.38
    LP 20.16 11.91 8.25 30.02 20.38 9.64 -0.54
    FT 23.64 14.56 9.08 32.11 21.70 10.41 -1.14
    GloGen 20.16 11.91 8.24 28.38 18.99 9.39 -0.41
    (d) Pre-trained: Sensors - Target: UCI
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 27.72 19.16 8.56 30.40 22.40 8.00 -0.70
    LP 27.39 19.05 8.34 30.27 22.11 8.16 -0.72
    FT 29.17 20.30 8.87 30.65 22.41 8.24 -0.88
    GloGen 26.58 18.53 8.05 29.40 21.54 7.86 -0.46
    10
    Scratch 27.78 18.65 9.13 30.31 22.32 7.98 -0.80
    LP 26.58 18.07 8.51 30.05 21.80 8.25 -0.52
    FT 28.80 20.65 8.15 28.85 21.43 7.42 -0.85
    GloGen 26.03 17.89 8.14 29.27 21.42 7.86 -0.40
    (e) Pre-trained: UCI - Target: BCG
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 27.10 15.20 11.90 37.37 24.84 12.52 -2.49
    LP 26.34 16.41 9.93 35.03 23.45 11.58 -1.98
    FT 31.40 20.75 10.64 34.49 24.01 10.48 -3.32
    GloGen 23.56 15.91 7.65 29.44 20.51 8.94 -0.96
    10
    Scratch 20.71 13.24 7.47 31.35 22.31 9.04 -0.41
    LP 22.33 13.05 9.28 32.65 21.45 11.20 -1.00
    FT 24.95 15.62 9.33 33.19 21.54 11.65 -1.92
    GloGen 20.17 12.54 7.63 30.40 20.99 9.40 -0.47
    (f) Pre-trained: UCI - Target: Sensors
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 37.20 26.16 11.04 33.73 24.45 9.28 -2.48
    LP 31.29 22.11 9.18 31.86 23.16 8.70 -1.42
    FT 33.90 24.46 9.43 31.84 23.67 8.17 -1.70
    GloGen 30.44 21.77 8.67 29.47 21.86 7.61 -1.07
    10
    Scratch 34.64 26.18 8.45 32.23 24.23 8.01 -1.67
    LP 29.62 21.09 8.52 30.54 22.37 8.17 -0.90
    FT 30.39 21.44 8.95 29.91 22.04 7.86 -1.12
    GloGen 29.22 20.80 8.42 28.50 20.87 7.63 -0.85
    We set the choice of which prompt to normalize and whether to apply
    clipping to the GloGen prompt as hyperparameters.
    3.4. Variance penalty
    The training set used to get a pre-trained model for PPG signals is
    typically imbalanced (Table 3). Consequently, the pre-trained model
    may perform optimally only for certain BP groups. The specificity
    towards a certain BP group have detrimental effects on the robustness
    of the model and gives biased knowledge when employing transfer
    learning with such a pre-trained model. To remedy this, we train
    GloGen with the regularization named Variance Penalty (VP).
    Let 𝑘 denote the index of the BP group, and let 𝑥𝑖,𝑘 denote the input
    PPG signal associated with BP group 𝑘. Suppose that 𝑥
    𝐼𝑃
    𝑖,𝑘 ∈ R𝐷, i.e., 𝑥
    𝐼𝑃
    𝑖,𝑘
    consists of 𝐷 time steps and 𝑥
    𝐼𝑃
    𝑖𝑗,𝑘 represents each time step. Then, the
    average of IP within BP group 𝑘 is given by
    ̄𝑥
    𝐼𝑃
    𝑗,𝑘 =
    1
    𝐺𝑘
    𝐺𝑘 ∑
    𝑖
    𝑥
    𝐼𝑃
    𝑖𝑗,𝑘, (2)
    where 𝐺𝑘
    stands for the number of PPG signals in BP group 𝑘. Then,
    VP is defined as follows:
    𝑉 𝑃 =
    1
    𝐷

    1
    𝐾
    𝐷∑
    𝑗
    𝐾∑
    𝑘
    (𝜇𝑗
    𝐼𝑃 − ̄𝑥
    𝐼𝑃
    𝑗,𝑘 )
    2
    , (3)
    where 𝜇𝑗
    𝐼𝑃 =
    1
    𝐾
    ∑𝐾
    𝑘
    ̄𝑥
    𝐼𝑃
    𝑗,𝑘 and K denotes the number of BP groups.
    Assuming that PPG signals within the same BP group are more similar
    to each other than those from different BP groups, we aim to generate
    distinctly different IPs for each BP group. By maximizing VP, we seek to
    enhance the learning process for underrepresented BP groups through
    IP and mitigate biased predictions for specific BP groups.
    However, using 𝐿𝑀𝑆𝐸 − 𝑉 𝑃 as the objective function to maximize
    VP results in highly unstable training, where 𝐿𝑀𝑆𝐸 denotes the Mean
    Table 2
    Hyper-parameter Range for Experiments. LR and WD denote
    learning rate and weight decay, respectively.
    Hyper-parameter range
    LR [1e−3, 1e−4, 1e-5]
    WD [0, 1e−3, 1e−4]
    𝜆 [1e−2 1.0,10.0]
    𝛾 [1e−2 1.0,10.0]
    𝑚 [1e−4, 1e−3, 1e−2, 1e−1, 1, 10]
    𝛼 [0.1 1 10 100 1000 10000]
    Normalization [𝑥
    𝐺𝑃
    , 𝑥
    𝐼𝑃
    , 𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛, Clipping]
    Squared Error (MSE). This is because the objective function decreases as
    VP increases, potentially resulting in no learning for prompt generation.
    Therefore, we use ReLU function that takes a zero value beyond a
    certain point and introduces a margin 𝑚 to control the extent of VP.
    𝑚 is a hyperparameter, and it adjusts the point at which ReLU function
    reaches zero. Thus, our objective function is as follows:
    𝐿 = 𝐿𝑀𝑆𝐸 + 𝛼 ⋅ 𝑅𝑒𝐿𝑈(𝑚 − 𝑉 𝑃 )), (4)
    where 𝛼 is a hyperparameter used to control the influence between MSE
    and VP. By minimizing Eq. (4), Prompt Generator ensures the diversity
    of IPs across BP groups and each IP transforms the input PPG signal
    to enhance the predictive performance of the pre-trained model with
    encouraging robustness across BP groups.
  4. Experiments
    4.1. Experimental setting
    We conduct a regression task to estimate SBP and DBP using PPG
    signals for BP prediction. For few-shot transfer learning, we reorganize
    three benchmark datasets introduced in [13]: BCG, Sensors, and UCI.
    Computers in Biology and Medicine 183 (2024) 109216
    4
    T. Kim et al.
    Fig. 2. Performance by BP group. From (a) to (f), each represents MAE within each BP group according to the pair of the dataset used for pre-training and the target task. We
    provide MAE for the Hypo, Normal, Prehyper, and Hyper2 groups, respectively. The Hypo and Hyper2 groups generally show higher MAE. We conduct all experiments for both
    5-shot and 10-shot settings. The unit for all values corresponding to the y-axis is mmHg.
    Specifically, we annotate each data instance according to its BP group
    based on SBP and DBP. Instances with excessively high or low SBP and
    DBP values are considered outliers and removed. BP groups are divided
    into hypotensive (Hypo), normotensive (Normal), prehypertensive (Pre￾hyper), and hypertensive (Hyper2) categories. The detailed criteria for
    defining these BP groups and the data distribution for each dataset can
    be found in Table 3.
    To select the best model, we divide BCG and Sensors datasets into
    5 folds, respectively, with 3 folds composing the training set in cross￾validation manner. Considering the size of UCI dataset, it is divided into
    3 folds to form the training–validation-test sets [13]. Each reorganized
    dataset is used to train the pre-trained model and conduct transfer
    learning with few-shot training set on the target task. The target task
    involves BP estimation for datasets different from those used to develop
    the pre-trained model. The few-shot training sets are composed of
    instances selected as 5-shot or 10-shot for each BP group from the
    training set on the target task. For model validation, we construct
    the 5-shot validation set from the validation fold of the target task.
    We evaluate the full shot of the test set for accurate performance
    evaluation. We adopt the ResNet1D model [23] as the architecture
    for all experiments to ensure a fair comparison of various transfer
    learning methods. ResNet1D mirrors the original ResNet [35] but uses
    1D convolutions instead of 2D, making it suitable for processing time￾series data while retaining the core design of residual blocks and skip
    connections. The model also includes Batch Normalization and ReLU
    activation. In the GloGen experiments, except in cases where PCA
    encoding is used, the encoder used to generate 𝑍𝑒𝑚𝑏𝑒𝑑 for the Prompt
    Generator is derived from the penultimate output of the ResNet1D pre￾trained model. For model selection, we use validation Group Average
    MAE. All hyperparameter searching is conducted using grid search, and
    the hyperparameter search space can be found in Table 2.
    4.2. Performance evaluation of GloGen
    Table 1 shows the results of experiments conducted on few-shot
    transfer learning. We compare GloGen with various methods of training
    Table 3
    Dataset Distribution. The first and second rows represent the criteria used for categoriz￾ing BP groups based on SBP and DBP, while the remaining rows present the number of
    instances for each group within each benchmark dataset [13,34]. We exclude instances
    that deviate from the criteria. When SBP and DBP satisfy the criteria for different BP
    groups, we assign the instance to the higher of the two BP groups.
    Hypo Normal Prehyper Hyper2 Total
    SBP [mmHg] [80∼90] [90∼120] [120∼140] [140∼180] –
    DBP [mmHg] [40∼60] [60∼80] [80∼90] [90∼120] –
    BCG 18 1466 1274 295 3053
    Sensors 78 2849 3890 4012 10 829
    UCI 3881 127 042 135 632 133 902 400 457
    MIMIC 45 954 539 814 312 394 171 919 1070081
    the target task: models trained from scratch (Scratch), models where
    only the last layer 𝑓𝜑
    𝑟𝑒𝑔 of the pre-trained model is randomly initialized
    and trained for the target task (LP), and models where the pre-trained
    model are fully fine-tuned for the target task (FT). We use Mean
    Absolute Error (MAE) as the evaluation metric. To evaluate the model
    robustness on BP groups as well as the average MAE, 𝐿𝐷𝑎𝑡𝑎, we also
    introduce group average MAE, 𝐿𝐺𝑟𝑜𝑢𝑝:
    𝐿𝐷𝑎𝑡𝑎 ∶=
    𝑁
    1
    𝑁∑
    𝑖=1
    |𝑓(𝑥𝑖
    ) − 𝑦𝑖
    |, (5)
    𝐿𝐺𝑟𝑜𝑢𝑝 ∶=
    1
    𝐾
    𝐾∑
    𝑘=1 𝐺
    1
    𝑘
    𝐺𝑘 ∑
    𝑖=1
    |𝑓(𝑥
    𝑘
    𝑖
    ) − 𝑦
    𝑘
    𝑖
    |, (6)
    where 𝑁, 𝐺𝑘
    , and 𝐾 denote the number of data in the test set, the
    number of test data belonging to the BP group, and the number of
    existing BP groups. Note that the 𝑥𝑖
    is the prompted PPG signal when
    we evaluate GloGen. While 𝐿𝐷𝑎𝑡𝑎 evaluates the overall BP estimation
    ability on the test distribution, it does not offer a clear explanation of
    the negative impacts caused by the bias of the model towards specific
    BP groups, given that the test distribution does not have a uniform
    distribution across BP groups. Conversely, 𝐿𝐺𝑟𝑜𝑢𝑝 is a evaluation metric
    Computers in Biology and Medicine 183 (2024) 109216
    5
    T. Kim et al.
    Fig. 3. Ablation Study on 𝛾 and 𝜆 with UCI dataset pre-training and Sensors Dataset
    as the target. N/A denote 𝛾 = 0 or 𝜆 = 0, i.e., no IP or no GP.
    for the predictive performance across BP groups with equal weighting,
    thereby serving as an indicator of the model’s robustness.
    Our objective is to estimate BP with better generalization ability on
    the test distribution and balanced predictive performance across dif￾ferent BP groups using the pre-trained model. We conduct experiments
    across every possible pre-trained — target pairs across three benchmark
    datasets. The results, elaborated in Table 1 show that GloGen aligns
    most closely with our goal. Notably, GloGen achieves the lowest total
    average MAE in most cases, consistently demonstrating its superior
    performance across both 5-shot and 10-shot settings. Moreover, GloGen
    exhibits a stronger correlation than the baselines, as reflected in the 2
    metric, which indicates the correlation between actual and estimated
    blood pressure values
    Moreover, the prompts in GloGen also contribute to the robustness
    of BP estimation. GloGen demonstrates the best total group average
    MAE in most cases, barring a few exceptions. Where it does not secure
    the top position, the robustness of GloGen remains on par with the
    best-performing baselines. It signifies that GloGen adeptly controls the
    trade-off between generalization and robustness with merely a few
    instances of the target task. This balance is facilitated by GP for learning
    shared PPG features and IP, along with VP for promoting diverse
    personalized prompts. We provide further analytical support for these
    findings in the following sections.
    4.3. Performance evaluation of BP estimation by BP group
    Fig. 2 shows the performance across BP groups within the test set.
    When referring to the Hypo and Hyper2 groups collectively as the
    High-risk Group (HG) and the Normal and Prehyper groups as the Low￾risk Group (LG), we observe that HG consistently exhibits higher MAE
    than LG in all cases. Generally, compared to other baselines, GloGen
    demonstrates performance that is comparable to or exceeds that of
    the LG. Simultaneously, it encourages learning for the HG, thereby
    exhibiting robustness across different BP groups. In most scenarios,
    GloGen ensures enhanced robustness for the group with the worst
    performance. While FT shows comparable results to GloGen in some
    metrics from Table 1, it does not simultaneously guarantee generaliza￾tion performance across the test distribution and robustness across BP
    groups. In the cases where FT performs well for HG in Fig. 2, MAE of LG
    increases. When showing low MAE for Hypo and Normal, performance
    deteriorates for Prehyper and Hyper2 in FT, e.g., in the 5-shot setting
    of Fig. 2(e). Contrarily, GloGen successfully manages this trade-off.
    Additionally, GloGen demonstrates superior performance in the 5-shot
    setting compared to the 10-shot setting. Thus, GloGen also proves to be
    particularly effective in scenarios where the dataset for the target task
    is inevitably comprised of a small number of data instances, such as in
    the field of medicine.
    Table 4
    Performance variations of GloGen with GP normalization (GP Norm), 𝑥
    𝐺𝑃 (GP), 𝑥
    𝐼𝑃
    (IP) and Variance Penalty (VP). The experiment results are based on models pre￾trained on the UCI dataset and fine-tuned with only 5-shot learning on the Sensors
    dataset.
    Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    Scratch 37.20 26.16 11.04 33.73 24.45 9.28 -2.48
    LP 31.29 22.11 9.18 31.86 23.16 8.70 -1.42
    FT 33.90 24.46 9.43 31.84 23.67 8.17 -1.70
    w/o GP Norm 33.40 23.78 9.62 31.26 23.32 7.94 -1.70
    w/o GP 34.14 25.22 8.92 31.97 24.06 7.91 -1.75
    w/o IP 32.72 23.73 8.99 31.54 23.63 7.90 -1.50
    w/o VP 33.84 24.92 8.92 30.90 23.43 7.46 -1.70
    GloGen 30.44 21.77 8.67 29.47 21.86 7.61 -1.07
    Table 5
    Performance Evaluation of Model pre-trained on MIMIC dataset. Avg SBP and Avg DBP
    represent the mean SBP MAE and mean DBP MAE, respectively, and Avg refers to the
    sum of these two. For GA SBP and GA DBP, the mean is first calculated for each BP
    group individually for both SBP and DPB, and then these values are averaged across
    the BP groups. GA refers to the sum of SBP GA and DBP GA. 2
    refers to the coefficient
    of determination. All units are mmHg.
    Pre-trained: MIMIC - Target: VitalDB
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 25.63 15.61 10.03 32.74 20.51 12.23 -0.48
    LP 25.69 15.74 9.94 32.06 20.58 11.48 -0.40
    FT 25.00 15.37 9.63 31.38 20.14 11.24 -0.32
    GloGen 23.41 14.21 9.20 30.69 19.53 11.17 -0.04
    10
    Scratch 24.04 14.94 9.10 33.88 22.31 11.57 -0.14
    LP 26.11 16.20 9.91 31.47 20.41 11.06 -0.46
    FT 26.58 16.07 10.51 31.38 19.92 11.46 -0.61
    GloGen 23.78 14.49 9.29 29.48 18.76 10.71 -0.08
    4.4. The role of GloGen prompt
    Even when SBP and DBP values are similar, the shapes of PPG sig￾nals are not identical due to a variety of factors, such as measurement
    environments or individual patient characteristics (Fig. 4(a)). GloGen
    Prompt assists the pre-trained model in successfully performing BP
    estimation by making the shape of unseen PPG signals resemble those
    in the training set with similar SBP and DBP. In Fig. 4(a, b), we observe
    significant increases in the similarity between PPG signals from the
    training set and the test set that have comparable SBP and DBP. It not
    only occurs among PPG signals within the same BP groups but also
    across different BP groups, indicating the effective operation of GP in
    learning shared features in the training set.
    However, it is still evident that PPG signals within the same BP
    group, exhibit greater similarity. It demonstrates that Prompt Generator
    successfully utilizes BP group information when generating diverse IPs.
    Fig. 4(c) demonstrates how IP utilizes information about BP groups to
    transform the PPG signal, facilitating more robust BP estimation. As dis￾cussed in previous section, pre-trained models generally perform better
    in LG than in HG. GloGen encourages diversity among IPs through the
    regularization by VP. It ensures that IP adequately learns from BP group
    information. In Fig. 4(c), IPs are distinctly separated into clusters for
    HG and LG, indicating that the mechanism by which IP modifies the
    input signal varies between HG and LG. Moreover, within HG, IPs are
    differentiated based on BP groups, Hypo and Hyper2. Consequently,
    Prompt Generator generates IPs characterized by different features
    according to BP groups, resulting in GloGen’s ability to transform the
    PPG signal for robust BP estimation.
    4.5. Ablation studies
    Fig. 3 illustrates how the size and presence of GP and IP impact
    the model’s Average MAE and Group Average MAE performance. In
    the experiment where the model is pre-trained on the UCI dataset
    and fine-tuned with only 5-shot learning on the Sensors dataset, we
    Computers in Biology and Medicine 183 (2024) 109216
    6
    T. Kim et al.
    Fig. 4. (a) Comparison of shape and cosine similarity between two PPG signals that are present in the training set and the test set, respectively, each having similar SBP and
    DBP values. (b) Changes in shape and cosine similarity following the addition of the corresponding GloGen prompt to the PPG signals presented in (a). (c) UMAP visualization of
    the distribution for IPs generated by Prompt Generator for the test set, following transfer learning to the target task UCI dataset in 10-shot setting using GloGen, with the model
    pre-trained on Sensors dataset.
    Table 6
    Performance evaluation of GloGen on classification. All metrics are
    reported as percentages.
    Shot Method Acc Group Acc Precision Recall F1-Score
    5
    Scratch 41.7 29.3 32.2 29.3 29.5
    LP 23.3 37.1 26.9 37.1 17.0
    FT 31.3 36.1 40.7 36.1 21.7
    GloGen 43.5 43.0 57.7 43.0 28.2
    10
    Scratch 42.9 26.1 24.3 26.1 23.0
    LP 16.7 34.1 22.2 34.1 15.1
    FT 17.7 34.9 22.2 34.9 15.1
    GloGen 42.3 40.8 30.0 40.8 26.9
    observe performance changes on the Sensors dataset by adjusting the
    hyperparameters 𝜆 and 𝛾, which determine the sizes of GP and IP.
    We also analyze scenarios where GP or IP is absent (N/A). GloGen
    achieves the best Average MAE and Group Average performance when
    both GP and IP are appropriately utilized. It demonstrates that both
    instance-specific information (IP) and common information across the
    dataset (GP) are necessary for robust prediction. Table 4 demonstrates
    that the presence or absence of GP normalization, 𝑥
    𝐺𝑃
    , 𝑋𝐼𝑃 , and VP
    significantly influences GloGen’s performance. It also highlights that
    they each play a critical role in enhancing both the performance and
    robustness of GloGen.
    4.6. Pre-trained on a large-scale dataset (MIMIC)
    With the advancement of Electronic Health Records, extensive med￾ical data collection has become possible, leading to the creation of
    large-scale datasets like the MIMIC dataset. Transfer learning aims to
    effectively utilize the knowledge from pre-trained models on different
    target datasets, making few-shot transfer learning powerful by not
    requiring a large-scale target dataset. Therefore, it is important to
    verify whether a transfer learning algorithm shows valid performance
    improvements when applied to a large-scale pre-trained model. We
    apply GloGen to a model pre-trained on MIMIC to evaluate whether
    GloGen can effectively facilitate the knowledge of the MIMIC pre￾trained model, enhancing the model’s generalization performance and
    robustness regarding BP groups. The target dataset for this experiment
    is VitalDB [34].
    In medical data experiments, the performance of the model can
    significantly change depending on whether the same patients exist
    between the train and test sets, potentially leading to information
    leakage about the subjects. MIMIC and VitalDB are non-overlapping
    datasets regarding patient information, and the training and validation
    sets for VitalDB are also set to avoid overlapping with the test set
    Table 7
    Performance of GloGen and all baselines varies with the number of peaks in PPG signals.

of Peaks Scratch LP FT GloGen

4 41.5 ± 2.1 25.2 ± 4.3 26.6 ± 5.5 35.4 ± 4.7
5 35.2 ± 4.4 19.3 ± 5.2 28.7 ± 10.2 7.4 ± 6.3
8 99.6 ± 2.3 82.9 ± 2.0 97.4 ± 0.8 77.8 ± 15.3
subjects [34]. Table 5 shows the results. Except for the average DBP
MAE in the 10-shot setting, GloGen outperforms the baseline across
all other metrics. Notably, it significantly improves Group Average
performance, indicating that GloGen remains effective even for large￾scale pre-trained models and successfully learns a robust model for
blood pressure group changes.
4.7. Classification
GloGen excels not only in regression tasks for SBP and DBP but
also in classification tasks that directly predict BP groups. To assess
GloGen’s effectiveness across various metrics, we include average accu￾racy, group accuracy (the average of accuracy in each group), precision,
recall, and F1-score in Table 6. We evaluate BP classification per￾formance on the BCG dataset in the few-shot setting using a model
pre-trained on the Sensors dataset. GloGen achieves the best perfor￾mance in all metrics except for average accuracy in the 10-shot setting,
demonstrating that the transformation of PPG signals with GloGen is
versatile and applicable to different tasks.
4.8. Analysis of failure cases
We observe that the number of peaks in PPG signals significantly
impacts model performance. Fig. 5 illustrates a portion of the BCG data,
while Table 7 presents the MAE values calculated when the model,
pre-trained on the UCI dataset and fine-tuned with the 5-shot BCG
dataset, estimates the data shown in Fig. 5. Except for the Fine-tuned
(FT) model, the other models demonstrate the best performance on the
input PPG signal in Fig. 5 with 5 peaks. Fig. 5(c) depicts data with 8
peaks, where all baselines, including GloGen, exhibit very high MAE
values.
4.9. Discussion
Numerous studies have proposed novel model architectures to pre￾dict SBP and DBP solely from PPG signals [4,9,22,23,36]. [13] com￾pares the average MAE results of three representative architectures –
ResNet1D [23], MLPBP [22], and SpectroResNet [4] – across various
Computers in Biology and Medicine 183 (2024) 109216
7
T. Kim et al.
Fig. 5. PPG signals from the BCG dataset categorized by the number of peaks.
datasets. Among these, ResNet1D is selected as the pre-trained model
in our study due to its overall superior performance. However, as
illustrated in Tables 1 and 5, the prediction model lacks robustness
in our setting where the pre-trained dataset and the target dataset
differ. Traditional transfer learning methods, such as FT and LP, could
be employed to address this issue, but their robustness is inconsistent
depending on the combination of the pre-trained and target distribu￾tions (Fig. 2). GloGen, on the other hand, consistently demonstrates
superior performance and robustness, irrespective of the distributional
or dataset size differences between the train and target datasets.
The primary objective of transfer learning is to ensure that knowl￾edge obtained during pre-training can be effectively applied to target
datasets with different distributions. While recent BP estimation studies
using PPG signals have explored transfer learning, most have presented
limited results by confining the transfer learning to the same dataset.
For instance, [8,9] separate pre-train and target datasets by patient
ID within the MIMIC-III dataset, requiring additional demographic
information, which restricts the applicability of the framework to spe￾cific datasets. Similarly, [37] proposes a transfer learning framework
confined within the MIMIC-II dataset. [38] introduces a contrastive
learning-based pre-training framework using large-scale datasets, pair￾ing high-quality and low-quality PPG signals. While this framework
presents results for downstream tasks with target datasets substantially
different from the pre-trained dataset, it relies on large-scale data for
pre-training and lacks quantitative evaluations of model robustness
across different BP ranges. While previous studies have focused on full
fine-tuning for the target dataset, our experimental results in the few￾shot setting, however, show that GloGen significantly enhances model
robustness with only a small number of samples in target distribution.
Additionally, the prompt-based approach, which is not constrained by
specific model architectures, further highlights the potential impacts of
our studies.
5. Conclusion
In this paper, we propose GloGen, a novel framework for few-shot
transfer learning on BP estimation using PPG signal. We explicitly
present the performance differences across BP groups for representative
transfer learning methods and demonstrate the effectiveness of prompt
learning using GloGen. Thorough GP for the shared features and IP for
the instance-specific features, GloGen successfully manages the trade￾off between BP estimation performance on the target distribution and
robustness across different BP groups. We also introduce VP to en￾courage Prompt Generator to generate diverse IPs. As a result, Prompt
Generator effectively extracts BP group information corresponding to
each PPG signal and IP facilitates robust BP estimation by transforming
the PPG signals. GloGen is not only applicable to PPG signals but can
also be extended to all other signal data, which we aim to explore in
future work.
CRediT authorship contribution statement
Taero Kim: Writing – review & editing, Writing – original draft, Vi￾sualization, Validation, Software, Methodology, Formal analysis, Con￾ceptualization. Hyeonjeong Lee: Writing – review & editing, Project
administration, Methodology, Investigation. Minseong Kim: Writing –
review & editing, Visualization, Software. Kwang-Yong Kim: Writing –
review & editing, Funding acquisition, Conceptualization. Kyu Hyung
Kim: Writing – review & editing, Funding acquisition, Conceptualiza￾tion. Kyungwoo Song: Writing – review & editing, Writing – orig￾inal draft, Project administration, Investigation, Funding acquisition,
Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing finan￾cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Acknowledgments
This work was supported by Electronics and Telecommunications
Research Institute (ETRI) grant funded by the Korean government.
[24ZD1140, Regional Industry ICT Convergence Technology Advance￾ment and Support Project in Daegu-GyeongBuk (Medical)], and sup￾ported by the National Research Foundation of Korea (NRF) grant
funded by the Korea government (MSIT) (RS-2024-00457216), and the
grant (22183MFDS431) from Ministry of Food and Drug Safety in 2024,
and the Industrial Technology Alchemist Project Program (No. RS-
2024-00419010) funded By the Ministry of Trade, Industry & Energy
(MOTIE, Korea).
References
[1] X. Teng, Y. Zhang, Continuous and noninvasive estimation of arterial blood
pressure using a photoplethysmographic approach, in: Proceedings of the 25th
Annual International Conference of the IEEE Engineering in Medicine and Biology
Society (IEEE Cat. No. 03CH37439), Vol. 4, IEEE, 2003, pp. 3153–3156.
[2] W. Shi, C. Zhou, Y. Zhang, K. Li, X. Ren, H. Liu, X. Ye, Hybrid mod￾eling on reconstitution of continuous arterial blood pressure using finger
photoplethysmography, Biomed. Signal Process. Control 85 (2023) 104972.
[3] H. Samimi, H.R. Dajani, A PPG-based calibration-free cuffless blood pressure
estimation method using cardiovascular dynamics, Sensors 23 (8) (2023) 4145.
[4] G. Slapničar, N. Mlakar, M. Luštrek, Blood pressure estimation from photo￾plethysmogram using a spectro-temporal deep neural network, Sensors 19 (15)
(2019) 3420.
[5] M. Kim, H. Lee, K.-Y. Kim, K.-H. Kim, Deep learning model for blood pres￾sure estimation from PPG signal, in: 2022 IEEE International Conference on
Metrology for Extended Reality, Artificial Intelligence and Neural Engineering
(MetroXRAINE), IEEE, 2022, pp. 1–5.
[6] A. Tazarv, M. Levorato, A deep learning approach to predict blood pressure
from ppg signals, in: 2021 43rd Annual International Conference of the IEEE
Engineering in Medicine & Biology Society, EMBC, IEEE, 2021, pp. 5658–5662.
Computers in Biology and Medicine 183 (2024) 109216
8
T. Kim et al.
[7] G. Frederick, et al., PPG signals for hypertension diagnosis: A novel method using
deep learning models, 2023, arXiv preprint arXiv:2304.06952.
[8] Y. Zhang, X. Ren, X. Liang, X. Ye, C. Zhou, A refined blood pressure estimation
model based on single channel photoplethysmography, IEEE J. Biomed. Health
Inf. 26 (12) (2022) 5907–5917.
[9] Z. Liu, Y. Zhang, C. Zhou, Bigru-attention for continuous blood pressure trends
estimation through single channel PPG, Comput. Biol. Med. 168 (2024) 107795.
[10] C.-T. Yen, S.-N. Chang, C.-H. Liao, Deep learning algorithm evaluation of hy￾pertension classification in less photoplethysmography signals conditions, Meas.
Control 54 (3–4) (2021) 439–445.
[11] E.-S.A. El-Dahshan, M.M. Bassiouni, S.K. Khare, R.-S. Tan, U.R. Acharya, Ex￾HyptNet: An explainable diagnosis of hypertension using EfficientNet with PPG
signals, Expert Syst. Appl. (2023) 122388.
[12] E. Martinez-Ríos, L. Montesinos, M. Alfaro-Ponce, A machine learning approach
for hypertension detection based on photoplethysmography and clinical data,
Comput. Biol. Med. 145 (2022) 105479.
[13] S. González, W.-T. Hsieh, T.P.-C. Chen, A benchmark for machine-learning based
non-invasive blood pressure estimation using photoplethysmogram, Sci. Data 10
(1) (2023) 149.
[14] L. Breiman, Random forests, Mach. Learn. 45 (2001) 5–32.
[15] S. Maqsood, S. Xu, M. Springer, R. Mohawesh, A benchmark study of ma￾chine learning for analysis of signal feature extraction techniques for blood
pressure estimation using photoplethysmography (PPG), Ieee Access 9 (2021)
138817–138833.
[16] K. Duan, Z. Qian, M. Atef, G. Wang, A feature exploration methodology for learn￾ing based cuffless blood pressure measurement using photoplethysmography, in:
2016 38th Annual International Conference of the IEEE Engineering in Medicine
and Biology Society, EMBC, IEEE, 2016, pp. 6385–6388.
[17] G. Wang, M. Atef, Y. Lian, Towards a continuous non-invasive cuffless blood
pressure monitoring system using PPG: Systems and circuits review, IEEE Circuits
Syst. Magaz. 18 (3) (2018) 6–26.
[18] T. Treebupachatsakul, A. Boosamalee, S. Shinnakerdchoke, S. Pechprasarn, N.
Thongpance, Cuff-less blood pressure prediction from ecg and ppg signals using
Fourier transformation and amplitude randomization preprocessing for context
aggregation network training, Biosensors 12 (3) (2022) 159.
[19] Y. Yoon, J.H. Cho, G. Yoon, Non-constrained blood pressure monitoring using
ECG and PPG for personal healthcare, J. Med. Syst. 33 (2009) 261–266.
[20] B.L. Hill, N. Rakocz, Á. Rudas, J.N. Chiang, S. Wang, I. Hofer, M. Cannesson,
E. Halperin, Imputation of the continuous arterial line blood pressure waveform
from non-invasive measurements using deep learning, Sci. Rep. 11 (1) (2021)
15755.
[21] N. Ibtehaz, S. Mahmud, M.E. Chowdhury, A. Khandakar, M. Salman Khan, M.A.
Ayari, A.M. Tahir, M.S. Rahman, PPG2ABP: Translating photoplethysmogram
(PPG) signals to arterial blood pressure (ABP) waveforms, Bioengineering 9 (11)
(2022) 692.
[22] B. Huang, W. Chen, C.-L. Lin, C.-F. Juang, J. Wang, MLP-BP: A novel framework
for cuffless blood pressure measurement with PPG and ECG signals based on
MLP-mixer neural networks, Biomed. Signal Process. Control 73 (2022) 103404.
[23] F. Schrumpf, P. Frenzel, C. Aust, G. Osterhoff, M. Fuchs, Assessment of non￾invasive blood pressure prediction from ppg and rppg signals using deep learning,
Sensors 21 (18) (2021) 6022.
[24] B. Lester, R. Al-Rfou, N. Constant, The power of scale for parameter-efficient
prompt tuning, 2021, arXiv preprint arXiv:2104.08691.
[25] X.L. Li, P. Liang, Prefix-tuning: Optimizing continuous prompts for generation,
2021, arXiv preprint arXiv:2101.00190.
[26] P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, G. Neubig, Pre-train, prompt,
and predict: A systematic survey of prompting methods in natural language
processing, ACM Comput. Surv. 55 (9) (2023) 1–35.
[27] X. Liu, K. Ji, Y. Fu, W.L. Tam, Z. Du, Z. Yang, J. Tang, P-tuning v2: Prompt
tuning can be comparable to fine-tuning universally across scales and tasks, 2021,
arXiv preprint arXiv:2110.07602.
[28] M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, S.-N. Lim,
Visual prompt tuning, in: European Conference on Computer Vision, Springer,
2022, pp. 709–727.
[29] C. Oh, H. Hwang, H.-y. Lee, Y. Lim, G. Jung, J. Jung, H. Choi, K. Song, Blackvip:
Black-box visual prompting for robust transfer learning, in: Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp.
24224–24235.
[30] H. Bahng, A. Jahanian, S. Sankaranarayanan, P. Isola, Exploring visual prompts
for adapting large-scale models, 2022, arXiv preprint arXiv:2203.17274.
[31] K. Zhou, J. Yang, C.C. Loy, Z. Liu, Conditional prompt learning for vision￾language models, in: Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, 2022, pp. 16816–16825.
[32] K. Zhou, J. Yang, C.C. Loy, Z. Liu, Learning to prompt for vision-language
models, Int. J. Comput. Vis. 130 (9) (2022) 2337–2348.
[33] H. Xue, F.D. Salim, Promptcast: A new prompt-based learning paradigm for time
series forecasting, IEEE Trans. Knowl. Data Eng. (2023).
[34] W. Wang, P. Mohseni, K.L. Kilgore, L. Najafizadeh, PulseDB: A large, cleaned
dataset based on MIMIC-III and VitalDB for benchmarking cuff-less blood
pressure estimation methods, Front. Digit. Health 4 (2023) 1090854.
[35] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
2016, pp. 770–778.
[36] D. Dai, Z. Ji, H. Wang, Non-invasive continuous blood pressure estimation from
single-channel PPG based on a temporal convolutional network integrated with
an attention mechanism, Appl. Sci. 14 (14) (2024) 6061.
[37] H.M. Koparır, Ö. Arslan, Cuffless blood pressure estimation from photoplethys￾mography using deep convolutional neural network and transfer learning,
Biomed. Signal Process. Control 93 (2024) 106194.
[38] C. Ding, Z. Guo, Z. Chen, R.J. Lee, C. Rudin, X. Hu, SiamQuality: a ConvNet￾based foundation model for photoplethysmography signals, Physiol. Meas. 45 (8)
(2024) 085004.
Computers in Biology and Medicine 183 (2024) 109216
9

Réflexion

Thought about GloGen for BP estimation,持续 几秒

Réponse

下面将对该论文进行较为全面和详细的介绍和总结,包括研究背景、问题动机、方法创新点、实验设置、结果分析和意义展望等方面。


论文题目:GloGen: PPG prompts for few-shot transfer learning in blood pressure estimation

作者和单位

  • 作者:Taero Kim、Hyeonjeong Lee、Minseong Kim、Kwang-Yong Kim、Kyu Hyung Kim、Kyungwoo Song
  • 所属机构:
    a. 韩国延世大学统计与数据科学系与应用统计学系
    b. 韩国电子与电信研究院(ETRI)大邱-庆北研究部

期刊信息:Computers in Biology and Medicine


研究背景与意义

血压(Blood Pressure, BP)监测是心血管健康管理的关键指标。传统的血压测量依赖袖带或有创动脉测压(Arterial Blood Pressure, ABP)方法,这些方法在实时、连续和无创监测方面存在诸多不便。而光电容积描记图(Photoplethysmogram, PPG)是一种可通过简单传感器非侵入性获取的生理信号,已被证实与ABP存在一定相关性,因此利用PPG来预测血压有望实现便捷、实时的无创连续监测。

近年来,随着深度学习的发展,不少研究已尝试利用PPG信号直接预测收缩压(SBP)和舒张压(DBP)。然而,即使在应用深度学习的前提下,模型在应对跨数据集或不同人群(如低血压和高血压高风险人群)时的泛化和鲁棒性仍然存在较大挑战。同时,为了进一步提高模型的实用性和适应性,迁移学习(transfer learning)成为焦点:在拥有大量数据和较高性能的预训练模型的基础上,将其能力迁移到新的、数据量有限的目标任务中。然而,在医疗数据领域中,目标任务的数据常常稀少(few-shot场景),这使得普通的迁移学习策略效用受限。

为了解决上述问题,本文提出了一种名为GloGen (Global Prompt and Prompt Generator)的新颖框架,旨在在极少数据(few-shot)条件下通过prompt式迁移学习手段,将预训练模型对PPG信号的知识高效传递给目标任务,从而提高血压估计的准确性和对高低血压等不同亚群体的鲁棒性。


问题动机与挑战

  1. 泛化能力不足:现有基于PPG的BP预测模型在面对与训练数据分布有较大差异的目标数据集时,性能明显下降,尤其对于血压过高或过低的高风险人群,精确预测尤为重要,但往往达不到理想的效果。

  2. 数据稀缺性:在医疗领域获得大量标注数据较为困难,许多场景中仅有极少的标注样本,使得传统深度模型或迁移学习方法难以有效泛化。

  3. 平衡整体性能与组间鲁棒性:模型在平均意义上的预测性能固然重要,但对于不同BP分组(如低血压、正常血压、前高血压、高血压人群)的预测一致性和稳定性同样关键。高风险组的预测不准确将导致更大的临床风险,因此实现对不同血压亚群的稳定预测成为一项重要任务。


方法论创新点:GloGen框架

GloGen的核心理念在于:使用prompt learning的思想来改造输入的PPG信号,使预训练模型更好地适应目标任务的分布和特性,从而在few-shot的场景下也能获得良好的表现。

  1. Global Prompt (GP)
    GP是一组可训练参数,与任意输入PPG信号线性叠加(加法)从而得到改造后的输入。GP在所有样本上共享,用于捕捉目标数据集的全局特征或公共模式,从而帮助模型更好适配总体分布。

  2. Instance-wise Prompt (IP) & Prompt Generator
    与GP不同,IP会为每个输入样本动态生成特定的“提示”。即通过将输入数据映射到一个特征嵌入空间后,使用Prompt Generator(一个含反卷积、归一化与激活层的解码结构)为每个输入实例生成独特的IP。IP可视为针对单个样本的精细调整,通过强调或抑制特定模式,使模型在个体层面适应目标分布。
    此外,引入触发向量(trigger vector)可为Prompt Generator补充额外的可训练信息,从而在少数据条件下仍能改善prompt生成的质量。

  3. 多重正则化与归一化策略
    为防止prompt值过大干扰原始PPG特征,论文对生成的prompts进行归一化和剪裁,使其范围与原始PPG信号分布相匹配,从而保持训练稳定性。

  4. Variance Penalty (VP)
    为了增强模型在不同BP子群体(血压分组)下的鲁棒性,论文设计了VP正则项,以鼓励在不同血压分组间生成的IP具备多样性。通过在损失函数中引入VP,模型会倾向于为不同的BP分组生成更有辨识度的prompt,这有助于提高高风险组数据的预测精度,改善全局鲁棒性。


实验设置与结果

  1. 数据集与实验场景
    实验使用文献中已建立的三类基准数据集(BCG、Sensors和UCI),同时也从MIMIC和VitalDB等大规模数据集中评估方法在迁移学习中的适用性。作者将BP值按范围分为低血压(Hypo)、正常血压(Normal)、前高血压(Prehyper)和高血压(Hyper2)四组,以衡量模型在不同血压区间的性能。
    实验采用few-shot设定,如每个BP组仅有5或10个训练样本。同时通过对比多种迁移策略(从零训练、仅微调最后一层、完整细粒度微调以及本文的GloGen)来评价方法有效性。

  2. 评价指标

    • 使用平均绝对误差(MAE)评价整体预测性能。
    • 使用分组平均误差(Group Average MAE)评估在各BP子群上性能的一致性和鲁棒性。
    • 使用决定系数R^2评估预测结果与真实血压值之间的相关性。
  3. 关键实验结果

    • GloGen在几乎所有测试场景下(不同预训练数据集与目标数据集的组合)均表现出领先的平均MAE和组平均MAE性能。
    • 相对于传统迁移学习(如fine-tuning或linear probing),GloGen不仅提升了整体精度,还有效降低了高风险组(Hypo与Hyper2)的预测误差。
    • 尤其在5-shot这样极端少样本学习条件下,GloGen的优越性更加明显。
    • 当在超大规模数据集(如MIMIC)预训练并转移到另一个数据集(VitalDB)时,GloGen同样显著提升模型对不同血压分组的鲁棒性。
  4. 消融实验与附加分析
    论文还讨论了不使用GP或IP、不同的正则化策略等设置对性能的影响。结果显示,没有其中任一模块(如GP、IP、VP)都会导致性能下降,表明各组件在GloGen框架中都起着关键作用。
    此外论文探讨了PPG信号自身特性(如波峰数目)等对模型性能的影响。


贡献与意义

  1. 新颖的Prompt Learning框架
    将Prompt Learning概念从自然语言处理扩展到PPG时间序列信号处理中,利用GloGen prompts作为高维数据的可训练“输入增强”,为该领域的迁移学习提供了一种新思路。

  2. 提升few-shot场景下的迁移能力
    在医疗领域获取大规模目标数据非常困难。GloGen在仅有极少数据时仍能有效提升模型的准确性和鲁棒性,显示了在实际医疗场景中部署BP预测系统的巨大潜力。

  3. 改善高风险人群预测精度
    通过鼓励prompt的多样性和差异化,GloGen对原本难以准确预测的低血压和高血压组群体表现出了显著改进。

  4. 通用性与可扩展性
    GloGen不限制具体的模型结构或数据类型,原则上可扩展到其他生理信号或时间序列预测问题中。


展望

未来工作可能包括:

  1. 将GloGen应用于其他生理数据(如ECG)或更多下游任务(如疾病分类、并发症预测)。
  2. 结合其他强化策略(如对抗训练、少样本数据增强)进一步提高鲁棒性。
  3. 优化Prompt生成网络的结构和正则化策略,以适应更多复杂的数据分布与预测任务。

总结
总体而言,本文提出的GloGen框架通过引入Global Prompt与Instance-wise Prompt的双Prompt策略、并结合Variance Penalty正则提升了迁移学习在PPG血压预测中的适应性与鲁棒性。实验结果显示,在仅有极少数样本的数据集上,GloGen仍能在整体预测精度与分组鲁棒性方面显著超越传统方法,为在真实医疗场景中实现高效、精准、可推广的无创血压估计提供了具有参考价值的新途径。

Question

这篇论文是大模型吗,详细说明,看不懂。结合文章的内容,详细介绍

Réflexion

Thought about large model and GloGen,持续 几秒

Réponse

这篇论文并不是在讨论“大模型”(Large Language Model)或者类似 ChatGPT 这样的超大规模预训练语言模型。相反,它是一篇关于生理信号处理和迁移学习方法的研究论文。该论文关注的是利用已有预训练模型(在大量数据上训练好的模型)来帮助解决血压(Blood Pressure, BP)估计的下游任务,尤其是当我们只有极少数目标数据(few-shot场景)时,如何通过一种叫做“Prompt Learning”的策略来提高模型的泛化能力和对不同血压人群(如低血压、高血压)的鲁棒性。

下面将以通俗和详细的方式介绍论文的核心思路和方法。

背景与问题

  1. 血压监测的重要性:连续、无创地监测血压对于心血管健康管理很重要。传统血压测量要用袖带或者有创测量血管压力,比较繁琐和不方便。因此,科学家们希望利用简单且非侵入性可采集的生理信号(如光电容积描记图PPG信号)来预测血压。这意味着只要戴一个小型传感器,就能在日常生活中连续监测血压,对健康管理有很大意义。

  2. 已有挑战

    • 当前用PPG直接预测血压的深度学习方法在面对新的数据集或者不同人群(例如高血压、低血压)时,性能往往会显著下降。模型对某些群体(高风险人群)的预测不准确,这在医疗应用中是个大问题。
    • 医疗数据获取困难导致目标任务数据量极少(few-shot),使得直接在目标数据上训练模型变得非常困难。
  3. 迁移学习的需求: 有了一个已经在大数据集上训练好的模型,我们希望将其知识迁移到一个数据规模很小、分布又不同的目标数据集上。这是迁移学习的目的。传统迁移学习方法(微调整个模型或只微调最后一层)在few-shot场景中往往效果不足够理想。

论文贡献与方法简介

这篇论文提出了一种名为 GloGen 的新方法。这个方法不是“大模型”,而是对输入信号进行一种“提示式”(prompt-based)的改造,让已经训练好的模型更好地适应新数据。

Prompt Learning是什么?

在自然语言处理(NLP)中,Prompt Learning是指不直接修改已训练好的大模型参数,而是通过加入一些额外的、可训练的提示(prompts)来让模型在新任务上表现更好。这类似于给模型输入做一些巧妙的“引导”,使得模型在面对新任务时更容易适应。

虽然该思想最初来自NLP的大型预训练语言模型,但这篇论文将Prompt Learning的理念移植到生理信号处理领域,用于PPG信号的血压估计。

GloGen框架的核心组成

  1. Global Prompt (GP)

    • 这是针对整个数据集统一使用的提示向量。它与每个输入PPG信号简单相加,帮助模型在所有目标数据上都进行相似的全局调整。
    • 可以理解为对输入数据做一个统一的“补偿”或“偏移”,从而让预训练模型对新数据的特性更敏感。
  2. Instance-wise Prompt (IP) & Prompt Generator

    • 除了全局提示,该方法还为每个输入样本动态生成专属的提示(IP)。
    • 如何生成IP?通过一个名为Prompt Generator的小网络,它以输入的特征嵌入(embedding)为基础,生成适用于当前样本的提示信号。这样,每个输入PPG数据都会得到一个“专属的修饰”。
    • IP相当于针对个体差异和具体血压分组特征进行的细粒度引导,帮助模型更好区分不同人群(例如低血压和高血压的形态差异)。
  3. Variance Penalty (VP)

    • 为了增强模型在不同血压人群中的鲁棒性,论文设计了一个方差惩罚项。此正则化项鼓励为不同血压分组生成的IP具有更大差异,使模型能更清晰地区分和适应不同血压组的特征。
    • 简单说,VP让模型不要只对一种血压类型最敏感,而是学会区分多个群组,从而在高风险人群(如高血压)上也获得较好表现。

实验与结果

作者在三个基准数据集(BCG、Sensors、UCI)和一个超大数据集(MIMIC)上进行实验,采用few-shot设定(如每个血压分组只有5或10条训练样本)。实验结果表明:

  • GloGen方法与传统方法(从零训练、仅微调最后一层、全面微调预训练模型)相比,能够在总体预测误差和不同血压分组上的预测一致性方面取得显著提升。
  • 特别是在高风险的低血压和高血压人群上,GloGen生成的个性化提示提高了模型对这些群组的预测精度和鲁棒性。
  • 当预训练模型来自大规模数据集(如MIMIC)时,GloGen仍能在新数据集上取得优秀表现,证明了方法的通用性。

总结

  • 这篇论文并不是在提出一个“大语言模型”或者“超大规模模型”,而是探索如何利用Prompt Learning的思路,让已有的PPG血压预测模型在极少数据和分布转移场景下仍能优异表现。
  • GloGen通过全局提示(GP)和个体化提示(IP)的结合,以及引入方差惩罚(VP)保证分组多样性,最终实现了在few-shot条件下的血压预测性能提升和鲁棒性增强。

如果将其与“大模型”类比,区别在于这里没有引入数百亿参数的巨型预训练模型,也不涉及自然语言处理的大规模语言模型。本文使用的模型结构(如ResNet1D)是常规规模的深度模型,所提出的贡献在于一种新颖的“提示生成与添加”策略,而非构建巨型模型本身。

总结来说,这是一篇关于利用Prompt Learning理念对生理信号(PPG)进行特定输入改造,从而提高血压预测迁移学习能力的论文,而不是讨论现代意义上“大语言模型”的研究。

Question

好的,结合latex公式,详细说明文章的思路。举一个具体的例子,详细说明论文的思路Contents lists available at ScienceDirect
Computers in Biology and Medicine
journal homepage: www.elsevier.com/locate/compbiomed
GloGen: PPG prompts for few-shot transfer learning in blood pressure
estimation
Taero Kim a
, Hyeonjeong Lee c
, Minseong Kim c
, Kwang-Yong Kim c
, Kyu Hyung Kim c
,
Kyungwoo Song a,b,∗
a Department of Statistics and Data Science, Yonsei University, 50, Yonsei-ro, Seodaemun-gu, 03722, Seoul, Republic of Korea
b Department of Applied Statistics, Yonsei University, 50, Yonsei-ro, Seodaemun-gu, 03722, Seoul, Republic of Korea
c Daegu-Gyeongbuk Research Division, Electronics and Telecommunications Research Institute (ETRI), 1, Techno sunhwan-ro 10-gil, Yuga-eup,
Dalseong-gun, 42994, Daegu, Republic of Korea
A R T I C L E I N F O
MSC:
41A05
41A10
65D05
65D17
Keywords:
Photoplethysmogram signal
Prompt learning
Few-shot learning
Transfer learning
Blood pressure estimation
A B S T R A C T
With the rapid advancements in machine learning, its applications in the medical field have garnered increasing
interest, particularly in non-invasive health monitoring methods. Blood pressure (BP) estimation using
Photoplethysmogram (PPG) signals presents a promising opportunity for real-time, continuous monitoring.
However, existing models often struggle with generalization, especially for high-risk groups like hypotension
and hypertension, where precise predictions are crucial. In this study, we propose Global Prompt and Prompt
Generator (GloGen), a robust few-shot transfer learning framework designed to improve BP estimation using
PPG signals. GloGen employs a dual-prompt learning approach, combining Global Prompt (GP) for capturing
shared features across signals and an Instance-wise Prompt (IP) for generating personalized prompts for each
signal. To enhance model robustness, we also introduce Variance Penalty (VP) that ensures diversity among
the generated prompts. Experimental results on benchmark datasets demonstrate that GloGen significantly
outperforms conventional methods, both in terms of accuracy and robustness, particularly in underrepresented
BP groups, even in scenarios with limited training data. GloGen thus stands out as an efficient solution for
real-time, non-invasive BP estimation, with great potential for use in healthcare settings where data is scarce
and diverse populations need to be accurately monitored.

  1. Introduction
    BP monitoring is a critical component in managing cardiovascular
    health, providing essential insights into heart function. Traditional BP
    measurement methods are effective but require cumbersome equip￾ment. It makes continuous BP monitoring an inconvenience for patients
    who require it. Considering that BP is a crucial metric for assessing
    individual health status, the implementation of real-time BP monitoring
    via wearable devices holds the potential to yield significant economic
    and healthcare benefits. As a result, BP estimation has spurred interest
    in developing more convenient and non-invasive techniques, leading to
    the exploration of utilizing Photoplethysmogram (PPG) signals for BP
    estimation.
    PPG signal is a biosignal that indicates the changes of degree of light
    absorption based on the blood volume change or the velocity of blood
    flow. Numerous studies have extensively investigated the correlation
    between PPG and Arterial Blood Pressure (ABP) signals that contain
    ∗ Corresponding author at: Department of Applied Statistics, Yonsei University, 50, Yonsei-ro, Seodaemun-gu, 03722, Seoul, Republic of Korea.
    E-mail addresses: [email protected] (T. Kim), [email protected] (H. Lee), [email protected] (M. Kim), [email protected] (K.-Y. Kim),
    [email protected] (K.H. Kim), [email protected] (K. Song).
    actual BP information [1,2]. Unlike ABP signals, PPG signals are mea￾sured non-invasively, allowing for blood pressure monitoring without
    temporal or spatial constraints, which holds significant potential for
    widespread impact. In pursuit of more precise BP estimation, there is
    a burgeoning body of research focused on leveraging deep learning
    techniques to either estimate systolic and diastolic blood pressure (SBP
    and DBP, respectively) or to generate ABP directly [3–7]. One of the
    challenges in BP estimation is that the model’s predictive performance
    varies significantly across subgroups, such as disease, ethnicity, and
    gender [8,9]. In particular, for high-risk groups like patients with
    hypotension or hypertension, where precise BP measurement is more
    critically required, the performance tends to be somewhat less sat￾isfactory compared to individuals with normal blood pressure [10–
    12].
    As vast amounts of data accumulate, the field of machine learning
    has demonstrated notable success with transfer learning. It involves
    https://doi.org/10.1016/j.compbiomed.2024.109216
    Received 31 March 2024; Received in revised form 21 September 2024; Accepted 25 September 2024
    Computers in Biology and Medicine 183 (2024) 109216
    Available online 8 October 2024
    0010-4825/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
    T. Kim et al.
    initially learning the general features of substantial data and then
    applying this knowledge to smaller datasets specific to a target task.
    Similarly, in the realms of medicine and healthcare, advancements in
    database systems like Electronic Health Records have enabled access
    to extensive medical datasets, such as MIMIC. Recently, these datasets
    have been used to evaluate diverse BP estimation algorithms using
    PPG, along with the construction of various benchmark datasets [13].
    However, research on the efficient transfer learning methods for BP
    estimation using these resources is not being actively conducted. Partic￾ularly, effectively applying models to target tasks requires specific data
    related to those tasks, but acquiring sufficient data for the target tasks
    is challenging due to the nature of medical data. Therefore, exploring
    desirable transfer learning methods that achieve better generalization
    ability for BP estimation and robustness across BP groups with a limited
    number of data is essential.
    In this study, we propose Global Prompt and Prompt Generator
    (GloGen), a novel few-shot transfer learning algorithm for BP esti￾mation using PPG signals. GloGen consists of Global Prompt (GP)
    composed of trainable parameters itself, and Instance-wise Prompt (IP)
    generated by Prompt Generator. Uniquely, GloGen combines two types
    of prompts into a PPG signal. This dual-prompt strategy allows our
    model to capture both common features across PPG signals and distinct
    features of individual PPG signals. Consequently, IP applies different
    augmentation policies for each instance, ensuring robust estimation
    by the model even for underrepresented BP groups in the pre-trained
    models. Moreover, we introduce Variance Penalty (VP) to encourage
    diversity among IPs. VP prevents the model from bias towards certain
    BP groups or generating similar IPs. By promoting the diverse prompts,
    we facilitate more personalized PPG signals, crucial for robust BP
    estimation especially in hypotensive or hypertensive BP groups.
    To ensure a fair evaluation of our method, we conduct experiments
    with reconstructing recently proposed benchmark datasets [13]. To
    evaluate the model robustness across different BP groups, we divide
    each instance according to specific criteria for assigning BP group and
    report both the performance for each group and the group average per￾formance, which is the average of these group-specific performances.
    It allows us to quantitatively assess whether our model performs BP
    estimation robustly across BP groups.
  2. Related works
    2.1. PPG-based blood pressure estimation
    PPG signals reflect the variations in the amount of light reflected
    due to differences in blood volume. One reason BP estimation based
    on PPG is actively studied is that PPG is a biometric signal that can be
    obtained non-invasively while showing a correlation with ABP signals,
    which are invasive [1,2]. Traditional ML methods like Random For￾est [14] extract features known to have a strong correlation with blood
    volume from PPG signals for BP estimation [15–17]. Additionally, some
    studies not only utilize PPG signals but also employ additional biosig￾nals containing other information on BP, such as Electrocardiogram
    (ECG) signals. [18,19]. There is also active research on generating ABP
    directly from PPG signals for continuous BP estimation [8,9,20,21].
    However, this study aims to directly estimate SBP and DBP using only
    PPG signals [4,22,23]. [13] evaluate various models for BP estimation
    through PPG signals and introduce various benchmark datasets for fair
    evaluation. Unlike previous studies, our research presents a transfer
    learning strategy that effectively utilizes the knowledge of a pre-trained
    model for target tasks. We also assume a few-shot setting where only a
    limited number of data are available for the target task. Moreover, in
    addition to the overall BP estimation performance, we introduce evalu￾ation metrics that reflect the differences in BP estimation performance
    across BP groups to assess the robustness of the model.
    2.2. Prompt learning
    Prompt learning, first introduced in NLP, serves as an alternative
    to the costly fine-tuning of pre-trained language models, utilizing a
    small number of additional parameters for model training [24,25].
    Prompt learning trains the optimal prompt to enable the model to adapt
    to the target task easily, demonstrating superior generalization per￾formance compared to fine-tuning [26,27]. In Computer Vision (CV),
    prompt learning has shown promising results in various recognition
    tasks by introducing trainable parameters and adding them to the
    input images [28,29]. By learning an optimal augmentation policy,
    prompt learning potentially restructures the input images to enable the
    model to achieve better generalization ability on the target task [30].
    Similarly, research in Vision–Language Models [31,32] and time-series
    forecasting [33] is gaining momentum. This study demonstrates that
    prompt learning is an effective transfer learning strategy for PPG signals
    as well. GloGen generates PPG prompts in a manner that ensures ro￾bustness to BP groups as well as overall model predictive performance,
    even with a limited amount of data for the target task. That is, PPG
    prompts reformulate the PPG signals to make models more robust. They
    can be integrated with input data in various ways, and we opt for a
    method directly adding the PPG prompt to the input data, ensuring it
    functions well regardless of the model structure [29].
    2.3. Comparison with existing research
    Existing research for BP estimation using PPG signal has generally
    progressed by improving model architectures [4,9,23]. However, our
    study operates differently by generating new prompts using only the
    input and output without making direct changes to the model itself.
    This approach implies that it is applicable to any model, as long as
    the model’s embeddings can be obtained. Additionally, some studies
    require demographic information [8,9], but GloGen only needs SBP and
    DBP from the training dataset, and the input PPG signal, as BP groups
    are determined by SBP and DBP. During testing, GloGen does not use
    group information and relies on the model’s predictions. Furthermore,
    our approach using prompt learning assumes a few-shot setting, unlike
    previous PPG-related transfer learning studies [8,9]. Specifically, we
    perform transfer learning with only 5 or 10 samples per blood pressure
    group, which contrasts with previous research that requires the entire
    training set of the target dataset for transfer learning. We also observe
    and improve performance differences across blood pressure groups de￾fined by small intervals. Many studies on BP estimation using PPG focus
    solely on average BP prediction performance [13,34]. However, we
    concentrate on enhancing model robustness as well, achieving this by
    incorporating minimal additional parameters through prompt learning.
  3. Methodology
    3.1. Global prompt and prompt generator
    We introduce two types of prompts: Global Prompt (GP) and
    Instance-wise Prompt (IP). GP is a set of trainable parameters itself and
    is equally added to all input PPG signals. On the other hand, IP reflects
    the individual characteristics of each input PPG signal. Therefore,
    each instance generates a different IP, for which we construct Prompt
    Generator. When estimating BP based on PPG signals, the common
    characteristics of all PPG signals are important for overall generaliza￾tion ability. In our methodology, GP is considered as a prompt that
    learns such shared features. However, for data belonging to a certain
    BP group, it is crucial to learn the specific characteristics of each data.
    IP addresses this need, enabling our model to robustly handle underrep￾resented BP groups such as hypotension or hypertension groups in the
    pre-trained model. Ultimately, to ensure robust estimations for different
    BP groups as well as overall predictive ability, we utilize both GP
    Computers in Biology and Medicine 183 (2024) 109216
    2
    T. Kim et al.
    Fig. 1. Overview of GloGen model. 𝑥
    𝐺𝑃 and 𝑥
    𝐼𝑃 denote Global Prompt (GP) and Instance-wise Prompt (IP), respectively. Encoder ℎ uses PPG signals to get embeddings 𝑍𝑒𝑚𝑏𝑒𝑑 for
    generating IP by Prompt Generator 𝑔(𝜙). PPG signal, combined with 𝑥
    𝐺𝑃 and 𝑥
    𝐼𝑃 , passes through the pre-trained model 𝑓 = 𝑓𝜃
    𝑝𝑒𝑛
    ◦𝑓𝜑
    𝑟𝑒𝑔 for BP estimation. 𝑉 𝑃 and 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 represent
    the variance of IPs across BP groups and the trigger vector, respectively.
    and IP. We name this dual-prompt strategy Global Prompt and Prompt
    Generator (GloGen). Fig. 1 shows the overview of GloGen.
    In GloGen, Prompt Generator 𝑔𝜙
    corresponds to the decoder within
    the encoder–decoder architecture. To generate optimal prompts, Glo￾Gen offers two configuration options. The first pertains to the configu￾ration of the encoder ℎ. By default, as an encoder in GloGen, we adopt
    a feature extractor 𝑓𝜃
    𝑝𝑒𝑛 that corresponds to all but the last layer of the
    pre-trained model 𝑓 = 𝑓𝜃
    𝑝𝑒𝑛
    ◦𝑓𝜑
    𝑟𝑒𝑔 , i.e., ℎ = 𝑓𝜃
    𝑝𝑒𝑛, where 𝑓𝜑
    𝑟𝑒𝑔 stands for
    the linear layer for regression. This approach aims to maximize the use
    of the pre-trained model’s knowledge for generating prompts. However,
    if 𝑓 is undertrained or strongly biased, it may lead to suboptimal
    embeddings 𝑍𝑒𝑚𝑏𝑒𝑑 for prompt generation. For these reasons, Principal
    Component Analysis (PCA) encoder 𝑓
    𝑝𝑐𝑎 can be a viable option, that
    is, ℎ = 𝑓
    𝑝𝑐𝑎. PCA encoder constructs a fixed projection matrix for the
    PCA space using few-shot data from the training set. This ensures that
    during the inference stage, the test set is projected onto this PCA space
    to obtain 𝑍𝑒𝑚𝑏𝑒𝑑 , thereby also ensuring faster inference times.
    The second option pertains to the construction of embeddings. The
    embedding 𝑍𝑒𝑚𝑏𝑒𝑑 , derived from the pre-trained model, extracts fea￾tures of the target task data by solely considering the characteristics of
    the data utilized during pre-training. Consequently, it fails to generate
    appropriate embeddings for information unique to the target task.
    Inspired by [29], we introduce a randomly initialized trainable vector,
    referred to as the trigger vector 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟. Unlike 𝑍𝑒𝑚𝑏𝑒𝑑 , 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 is not
    instance-specific. For a target task, a single 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 is concatenated
    uniformly across all 𝑍𝑒𝑚𝑏𝑒𝑑 and then input into Prompt Generator.
    𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 is updated during the training stage and models the input
    vector for Prompt Generator that represents the embedding for the
    additional information to generate the appropriate prompt adopting
    models to the target task. Note that the role of GP and 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 differ.
    While GP aims to learn the shared characteristics across all the training
    data, 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟, in conjunction with instance-specific information 𝑍𝑒𝑚𝑏𝑒𝑑 ,
    strives to facilitate the generation of more optimal prompts.
    3.2. Construction of GloGen prompts
    Let 𝑖 denote the data index. Both GP, 𝑥
    𝐺𝑃 , and IP, 𝑥
    𝐼𝑃
    𝑖
    , share
    the same dimension as the input PPG signal, 𝑥𝑖
    . 𝑥
    𝐺𝑃 is a learnable
    parameter added uniformly to all 𝑥𝑖
    . Through this process, 𝑥
    𝐺𝑃 is
    expected to capture common features within the training data. Unlike
    𝑥
    𝐺𝑃
    , 𝑥
    𝐼𝑃
    𝑖
    is generated by Prompt Generator 𝑔𝜙
    :
    𝑥
    𝐼𝑃
    𝑖
    = 𝑔𝜙◦ℎ(𝑥𝑖
    ) = 𝑔𝜙
    (𝑧𝑖
    ), (1)
    where 𝑧𝑖 ∈ 𝑍𝑒𝑚𝑏𝑒𝑑 in the default setting and 𝑧𝑖 ∈ [𝑍𝑒𝑚𝑏𝑒𝑑 , 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟]
    when using trigger vector. 𝑔𝜙
    consists of a set of deconvolution layers
    (also known as Transposed Convolution layers), batch normalization,
    and ReLU activation function. IP, in contrast to GP, learns the unique
    characteristics that are associated with individual data or groups of
    data. Therefore, IP regulates each instance through corresponding per￾sonalized prompt, imposing additional constraints on challenging BP
    groups that are underrepresented in the model. We construct GloGen
    prompts by using GP and IP. Finally, the input adding GloGen prompt
    is:
    𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖
    = 𝑥𝑖 + 𝜆 ⋅ 𝑥
    𝐺𝑃 + 𝛾 ⋅ 𝑥
    𝐼𝑃
    𝑖
    ,
    where 𝜆 and 𝛾 stand for the coefficients controlling the effects of GP
    and IP, respectively.
    𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖
    serves as the input to the pre-trained model, which then
    predicts SBP and DBP for BP estimation. To train GP and Prompt
    Generator, we use Mean Squared Error (MSE) as a loss function:
    𝐿𝑀𝑆𝐸 =
    1
    𝑁
    𝑁∑
    𝑖
    (𝑓(𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖
    ) − 𝑦𝑖
    )
    2
    ,
    where 𝑁 is the number of training data and 𝑦𝑖
    is ground-truth SBP and
    DBP corresponding to 𝑥𝑖
    . Note that we always freeze 𝑓𝜃
    𝑝𝑒𝑛. Concretely,
    in the default setting, we only update the 𝑔𝜙
    , GP, and 𝑓𝜑
    𝑟𝑒𝑔 for recon￾figuration on the embeddings. If we use 𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 or 𝑓
    𝑝𝑐𝑎, we also update
    𝑍𝑡𝑟𝑖𝑔𝑔𝑒𝑟 or use training data for the target task to get the projection
    matrix, respectively.
    3.3. Normalization
    As 𝑥
    𝐺𝑃 and 𝑥
    𝐼𝑃
    𝑖
    are trainable, they can take any value. If the value
    of the prompt is much larger than the actual PPG input, it overshadows
    the information of the original PPG input, leading to training instability
    or predictions. Therefore, we introduce the normalization using the
    statistics of the training dataset.
    Let 𝑥𝑖𝑗 denote 𝑗th component of 𝑥𝑖
    , 𝑥
    𝑃
    𝑖
    denote prompt vector, and
    𝑋𝑡𝑟 denote a set of input PPG signals in the training set. 𝑥
    𝑃
    can be any
    𝑥
    𝐺𝑃
    , 𝑥
    𝐼𝑃
    𝑖
    or 𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖
    . When max(𝑋𝑡𝑟) = max𝑖,𝑗 𝑥𝑖𝑗 and min(𝑋𝑡𝑟) = min𝑖,𝑗 𝑥𝑖𝑗
    denote the maximum and minimum value in 𝑋𝑡𝑟, respectively, the
    normalized 𝑥
    𝑃
    𝑖
    is:
    ̂𝑥
    𝑃
    𝑖
    =
    max(𝑋𝑡𝑟) − min(𝑋𝑡𝑟)
    max(𝑥
    𝑃
    𝑖
    ) − min(𝑥
    𝑃
    𝑖
    )
    (𝑥
    𝑃
    𝑖
    − min(𝑥
    𝑃
    𝑖
    )) + min(𝑋𝑡𝑟),
    where max(𝑥
    𝑃
    𝑖
    ) = max𝑗 𝑥
    𝑃
    𝑖𝑗 and min(𝑥
    𝑃
    𝑖
    ) = min𝑗 𝑥
    𝑃
    𝑖𝑗. By normalizing 𝑥
    𝑃
    𝑖
    and adding ̂𝑥
    𝑃
    𝑖
    to the original PPG, we aim to stabilize the training
    process and ensure more consistent predictions.
    Inspired by [29], we also consider clipping the GloGen prompt. This
    approach, akin to normalization, is known to enhance the stability of
    prompt learning. Clipping ensures that the GloGen prompt values are
    constrained within the range defined by min(𝑋𝑡𝑟) and max(𝑋𝑡𝑟):
    ̂𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 =





    min(𝑋𝑡𝑟) if 𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 < min(𝑋𝑡𝑟)
    𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 if min(𝑋𝑡𝑟) ≤ 𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 ≤ max(𝑋𝑡𝑟)
    max(𝑋𝑡𝑟) if max(𝑋𝑡𝑟) < 𝑥𝐺𝑙𝑜𝐺𝑒𝑛
    𝑖𝑗 .
    Computers in Biology and Medicine 183 (2024) 109216
    3
    T. Kim et al.
    Table 1
    Performance Evaluation of GloGen. From (a) to (f), each represents the performance according to the pair of the data used for pre-training and the target task data. We report
    average MAE and group average MAE for both SBP and DBP, respectively. The total average MAE (Avg) and total group average MAE (GA) refer to the sum of the respective
    metrics for SBP and DBP. We conduct all experiments for both 5-shot and 10-shot settings. In all experiments, we select the model based on the best validation total average MAE.
    2
    refers to the coefficient of determination. All units are mmHg.
    (a) Pre-trained: BCG - Target: Sensors
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 46.35 35.69 10.66 37.25 28.77 8.49 -4.11
    LP 38.94 28.43 10.51 34.79 25.60 9.19 -2.77
    FT 33.32 24.41 8.92 31.56 23.52 8.03 -1.55
    GloGen 32.47 23.70 8.77 30.20 22.84 7.36 -1.42
    10
    Scratch 35.81 26.11 9.70 32.75 24.12 8.62 -2.02
    LP 33.06 24.07 8.99 31.26 23.46 7.80 -1.54
    FT 32.72 24.25 8.46 30.78 23.21 7.57 -1.39
    GloGen 31.71 22.80 8.91 29.76 22.35 7.40 -1.35
    (b) Pre-trained: BCG - Target: UCI
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 32.65 22.08 10.57 33.68 23.69 9.99 -1.74
    LP 28.89 19.18 9.71 32.09 23.16 8.94 -1.02
    FT 28.19 19.15 9.04 32.10 22.73 9.37 -0.77
    GloGen 27.75 19.09 8.67 30.23 22.29 7.94 -0.72
    10
    Scratch 30.03 20.78 9.25 31.21 22.84 8.38 -1.16
    LP 26.09 17.37 8.72 31.37 22.72 8.64 -0.45
    FT 27.58 18.81 8.77 31.84 22.57 9.27 -0.64
    GloGen 25.74 17.63 8.11 30.55 22.39 8.16 -0.30
    (c) Pre-trained: Sensors - Target: BCG
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 25.35 14.23 11.13 34.26 22.45 11.82 -1.75
    LP 22.06 12.58 9.48 31.62 20.40 11.22 -0.94
    FT 30.28 18.95 11.34 42.02 28.32 13.70 -3.00
    GloGen 19.97 11.87 8.10 27.92 17.98 9.94 -0.30
    10
    Scratch 20.09 12.56 7.53 32.26 22.56 9.70 -0.38
    LP 20.16 11.91 8.25 30.02 20.38 9.64 -0.54
    FT 23.64 14.56 9.08 32.11 21.70 10.41 -1.14
    GloGen 20.16 11.91 8.24 28.38 18.99 9.39 -0.41
    (d) Pre-trained: Sensors - Target: UCI
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 27.72 19.16 8.56 30.40 22.40 8.00 -0.70
    LP 27.39 19.05 8.34 30.27 22.11 8.16 -0.72
    FT 29.17 20.30 8.87 30.65 22.41 8.24 -0.88
    GloGen 26.58 18.53 8.05 29.40 21.54 7.86 -0.46
    10
    Scratch 27.78 18.65 9.13 30.31 22.32 7.98 -0.80
    LP 26.58 18.07 8.51 30.05 21.80 8.25 -0.52
    FT 28.80 20.65 8.15 28.85 21.43 7.42 -0.85
    GloGen 26.03 17.89 8.14 29.27 21.42 7.86 -0.40
    (e) Pre-trained: UCI - Target: BCG
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 27.10 15.20 11.90 37.37 24.84 12.52 -2.49
    LP 26.34 16.41 9.93 35.03 23.45 11.58 -1.98
    FT 31.40 20.75 10.64 34.49 24.01 10.48 -3.32
    GloGen 23.56 15.91 7.65 29.44 20.51 8.94 -0.96
    10
    Scratch 20.71 13.24 7.47 31.35 22.31 9.04 -0.41
    LP 22.33 13.05 9.28 32.65 21.45 11.20 -1.00
    FT 24.95 15.62 9.33 33.19 21.54 11.65 -1.92
    GloGen 20.17 12.54 7.63 30.40 20.99 9.40 -0.47
    (f) Pre-trained: UCI - Target: Sensors
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 37.20 26.16 11.04 33.73 24.45 9.28 -2.48
    LP 31.29 22.11 9.18 31.86 23.16 8.70 -1.42
    FT 33.90 24.46 9.43 31.84 23.67 8.17 -1.70
    GloGen 30.44 21.77 8.67 29.47 21.86 7.61 -1.07
    10
    Scratch 34.64 26.18 8.45 32.23 24.23 8.01 -1.67
    LP 29.62 21.09 8.52 30.54 22.37 8.17 -0.90
    FT 30.39 21.44 8.95 29.91 22.04 7.86 -1.12
    GloGen 29.22 20.80 8.42 28.50 20.87 7.63 -0.85
    We set the choice of which prompt to normalize and whether to apply
    clipping to the GloGen prompt as hyperparameters.
    3.4. Variance penalty
    The training set used to get a pre-trained model for PPG signals is
    typically imbalanced (Table 3). Consequently, the pre-trained model
    may perform optimally only for certain BP groups. The specificity
    towards a certain BP group have detrimental effects on the robustness
    of the model and gives biased knowledge when employing transfer
    learning with such a pre-trained model. To remedy this, we train
    GloGen with the regularization named Variance Penalty (VP).
    Let 𝑘 denote the index of the BP group, and let 𝑥𝑖,𝑘 denote the input
    PPG signal associated with BP group 𝑘. Suppose that 𝑥
    𝐼𝑃
    𝑖,𝑘 ∈ R𝐷, i.e., 𝑥
    𝐼𝑃
    𝑖,𝑘
    consists of 𝐷 time steps and 𝑥
    𝐼𝑃
    𝑖𝑗,𝑘 represents each time step. Then, the
    average of IP within BP group 𝑘 is given by
    ̄𝑥
    𝐼𝑃
    𝑗,𝑘 =
    1
    𝐺𝑘
    𝐺𝑘 ∑
    𝑖
    𝑥
    𝐼𝑃
    𝑖𝑗,𝑘, (2)
    where 𝐺𝑘
    stands for the number of PPG signals in BP group 𝑘. Then,
    VP is defined as follows:
    𝑉 𝑃 =
    1
    𝐷

    1
    𝐾
    𝐷∑
    𝑗
    𝐾∑
    𝑘
    (𝜇𝑗
    𝐼𝑃 − ̄𝑥
    𝐼𝑃
    𝑗,𝑘 )
    2
    , (3)
    where 𝜇𝑗
    𝐼𝑃 =
    1
    𝐾
    ∑𝐾
    𝑘
    ̄𝑥
    𝐼𝑃
    𝑗,𝑘 and K denotes the number of BP groups.
    Assuming that PPG signals within the same BP group are more similar
    to each other than those from different BP groups, we aim to generate
    distinctly different IPs for each BP group. By maximizing VP, we seek to
    enhance the learning process for underrepresented BP groups through
    IP and mitigate biased predictions for specific BP groups.
    However, using 𝐿𝑀𝑆𝐸 − 𝑉 𝑃 as the objective function to maximize
    VP results in highly unstable training, where 𝐿𝑀𝑆𝐸 denotes the Mean
    Table 2
    Hyper-parameter Range for Experiments. LR and WD denote
    learning rate and weight decay, respectively.
    Hyper-parameter range
    LR [1e−3, 1e−4, 1e-5]
    WD [0, 1e−3, 1e−4]
    𝜆 [1e−2 1.0,10.0]
    𝛾 [1e−2 1.0,10.0]
    𝑚 [1e−4, 1e−3, 1e−2, 1e−1, 1, 10]
    𝛼 [0.1 1 10 100 1000 10000]
    Normalization [𝑥
    𝐺𝑃
    , 𝑥
    𝐼𝑃
    , 𝑥
    𝐺𝑙𝑜𝐺𝑒𝑛, Clipping]
    Squared Error (MSE). This is because the objective function decreases as
    VP increases, potentially resulting in no learning for prompt generation.
    Therefore, we use ReLU function that takes a zero value beyond a
    certain point and introduces a margin 𝑚 to control the extent of VP.
    𝑚 is a hyperparameter, and it adjusts the point at which ReLU function
    reaches zero. Thus, our objective function is as follows:
    𝐿 = 𝐿𝑀𝑆𝐸 + 𝛼 ⋅ 𝑅𝑒𝐿𝑈(𝑚 − 𝑉 𝑃 )), (4)
    where 𝛼 is a hyperparameter used to control the influence between MSE
    and VP. By minimizing Eq. (4), Prompt Generator ensures the diversity
    of IPs across BP groups and each IP transforms the input PPG signal
    to enhance the predictive performance of the pre-trained model with
    encouraging robustness across BP groups.
  4. Experiments
    4.1. Experimental setting
    We conduct a regression task to estimate SBP and DBP using PPG
    signals for BP prediction. For few-shot transfer learning, we reorganize
    three benchmark datasets introduced in [13]: BCG, Sensors, and UCI.
    Computers in Biology and Medicine 183 (2024) 109216
    4
    T. Kim et al.
    Fig. 2. Performance by BP group. From (a) to (f), each represents MAE within each BP group according to the pair of the dataset used for pre-training and the target task. We
    provide MAE for the Hypo, Normal, Prehyper, and Hyper2 groups, respectively. The Hypo and Hyper2 groups generally show higher MAE. We conduct all experiments for both
    5-shot and 10-shot settings. The unit for all values corresponding to the y-axis is mmHg.
    Specifically, we annotate each data instance according to its BP group
    based on SBP and DBP. Instances with excessively high or low SBP and
    DBP values are considered outliers and removed. BP groups are divided
    into hypotensive (Hypo), normotensive (Normal), prehypertensive (Pre￾hyper), and hypertensive (Hyper2) categories. The detailed criteria for
    defining these BP groups and the data distribution for each dataset can
    be found in Table 3.
    To select the best model, we divide BCG and Sensors datasets into
    5 folds, respectively, with 3 folds composing the training set in cross￾validation manner. Considering the size of UCI dataset, it is divided into
    3 folds to form the training–validation-test sets [13]. Each reorganized
    dataset is used to train the pre-trained model and conduct transfer
    learning with few-shot training set on the target task. The target task
    involves BP estimation for datasets different from those used to develop
    the pre-trained model. The few-shot training sets are composed of
    instances selected as 5-shot or 10-shot for each BP group from the
    training set on the target task. For model validation, we construct
    the 5-shot validation set from the validation fold of the target task.
    We evaluate the full shot of the test set for accurate performance
    evaluation. We adopt the ResNet1D model [23] as the architecture
    for all experiments to ensure a fair comparison of various transfer
    learning methods. ResNet1D mirrors the original ResNet [35] but uses
    1D convolutions instead of 2D, making it suitable for processing time￾series data while retaining the core design of residual blocks and skip
    connections. The model also includes Batch Normalization and ReLU
    activation. In the GloGen experiments, except in cases where PCA
    encoding is used, the encoder used to generate 𝑍𝑒𝑚𝑏𝑒𝑑 for the Prompt
    Generator is derived from the penultimate output of the ResNet1D pre￾trained model. For model selection, we use validation Group Average
    MAE. All hyperparameter searching is conducted using grid search, and
    the hyperparameter search space can be found in Table 2.
    4.2. Performance evaluation of GloGen
    Table 1 shows the results of experiments conducted on few-shot
    transfer learning. We compare GloGen with various methods of training
    Table 3
    Dataset Distribution. The first and second rows represent the criteria used for categoriz￾ing BP groups based on SBP and DBP, while the remaining rows present the number of
    instances for each group within each benchmark dataset [13,34]. We exclude instances
    that deviate from the criteria. When SBP and DBP satisfy the criteria for different BP
    groups, we assign the instance to the higher of the two BP groups.
    Hypo Normal Prehyper Hyper2 Total
    SBP [mmHg] [80∼90] [90∼120] [120∼140] [140∼180] –
    DBP [mmHg] [40∼60] [60∼80] [80∼90] [90∼120] –
    BCG 18 1466 1274 295 3053
    Sensors 78 2849 3890 4012 10 829
    UCI 3881 127 042 135 632 133 902 400 457
    MIMIC 45 954 539 814 312 394 171 919 1070081
    the target task: models trained from scratch (Scratch), models where
    only the last layer 𝑓𝜑
    𝑟𝑒𝑔 of the pre-trained model is randomly initialized
    and trained for the target task (LP), and models where the pre-trained
    model are fully fine-tuned for the target task (FT). We use Mean
    Absolute Error (MAE) as the evaluation metric. To evaluate the model
    robustness on BP groups as well as the average MAE, 𝐿𝐷𝑎𝑡𝑎, we also
    introduce group average MAE, 𝐿𝐺𝑟𝑜𝑢𝑝:
    𝐿𝐷𝑎𝑡𝑎 ∶=
    𝑁
    1
    𝑁∑
    𝑖=1
    |𝑓(𝑥𝑖
    ) − 𝑦𝑖
    |, (5)
    𝐿𝐺𝑟𝑜𝑢𝑝 ∶=
    1
    𝐾
    𝐾∑
    𝑘=1 𝐺
    1
    𝑘
    𝐺𝑘 ∑
    𝑖=1
    |𝑓(𝑥
    𝑘
    𝑖
    ) − 𝑦
    𝑘
    𝑖
    |, (6)
    where 𝑁, 𝐺𝑘
    , and 𝐾 denote the number of data in the test set, the
    number of test data belonging to the BP group, and the number of
    existing BP groups. Note that the 𝑥𝑖
    is the prompted PPG signal when
    we evaluate GloGen. While 𝐿𝐷𝑎𝑡𝑎 evaluates the overall BP estimation
    ability on the test distribution, it does not offer a clear explanation of
    the negative impacts caused by the bias of the model towards specific
    BP groups, given that the test distribution does not have a uniform
    distribution across BP groups. Conversely, 𝐿𝐺𝑟𝑜𝑢𝑝 is a evaluation metric
    Computers in Biology and Medicine 183 (2024) 109216
    5
    T. Kim et al.
    Fig. 3. Ablation Study on 𝛾 and 𝜆 with UCI dataset pre-training and Sensors Dataset
    as the target. N/A denote 𝛾 = 0 or 𝜆 = 0, i.e., no IP or no GP.
    for the predictive performance across BP groups with equal weighting,
    thereby serving as an indicator of the model’s robustness.
    Our objective is to estimate BP with better generalization ability on
    the test distribution and balanced predictive performance across dif￾ferent BP groups using the pre-trained model. We conduct experiments
    across every possible pre-trained — target pairs across three benchmark
    datasets. The results, elaborated in Table 1 show that GloGen aligns
    most closely with our goal. Notably, GloGen achieves the lowest total
    average MAE in most cases, consistently demonstrating its superior
    performance across both 5-shot and 10-shot settings. Moreover, GloGen
    exhibits a stronger correlation than the baselines, as reflected in the 2
    metric, which indicates the correlation between actual and estimated
    blood pressure values
    Moreover, the prompts in GloGen also contribute to the robustness
    of BP estimation. GloGen demonstrates the best total group average
    MAE in most cases, barring a few exceptions. Where it does not secure
    the top position, the robustness of GloGen remains on par with the
    best-performing baselines. It signifies that GloGen adeptly controls the
    trade-off between generalization and robustness with merely a few
    instances of the target task. This balance is facilitated by GP for learning
    shared PPG features and IP, along with VP for promoting diverse
    personalized prompts. We provide further analytical support for these
    findings in the following sections.
    4.3. Performance evaluation of BP estimation by BP group
    Fig. 2 shows the performance across BP groups within the test set.
    When referring to the Hypo and Hyper2 groups collectively as the
    High-risk Group (HG) and the Normal and Prehyper groups as the Low￾risk Group (LG), we observe that HG consistently exhibits higher MAE
    than LG in all cases. Generally, compared to other baselines, GloGen
    demonstrates performance that is comparable to or exceeds that of
    the LG. Simultaneously, it encourages learning for the HG, thereby
    exhibiting robustness across different BP groups. In most scenarios,
    GloGen ensures enhanced robustness for the group with the worst
    performance. While FT shows comparable results to GloGen in some
    metrics from Table 1, it does not simultaneously guarantee generaliza￾tion performance across the test distribution and robustness across BP
    groups. In the cases where FT performs well for HG in Fig. 2, MAE of LG
    increases. When showing low MAE for Hypo and Normal, performance
    deteriorates for Prehyper and Hyper2 in FT, e.g., in the 5-shot setting
    of Fig. 2(e). Contrarily, GloGen successfully manages this trade-off.
    Additionally, GloGen demonstrates superior performance in the 5-shot
    setting compared to the 10-shot setting. Thus, GloGen also proves to be
    particularly effective in scenarios where the dataset for the target task
    is inevitably comprised of a small number of data instances, such as in
    the field of medicine.
    Table 4
    Performance variations of GloGen with GP normalization (GP Norm), 𝑥
    𝐺𝑃 (GP), 𝑥
    𝐼𝑃
    (IP) and Variance Penalty (VP). The experiment results are based on models pre￾trained on the UCI dataset and fine-tuned with only 5-shot learning on the Sensors
    dataset.
    Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    Scratch 37.20 26.16 11.04 33.73 24.45 9.28 -2.48
    LP 31.29 22.11 9.18 31.86 23.16 8.70 -1.42
    FT 33.90 24.46 9.43 31.84 23.67 8.17 -1.70
    w/o GP Norm 33.40 23.78 9.62 31.26 23.32 7.94 -1.70
    w/o GP 34.14 25.22 8.92 31.97 24.06 7.91 -1.75
    w/o IP 32.72 23.73 8.99 31.54 23.63 7.90 -1.50
    w/o VP 33.84 24.92 8.92 30.90 23.43 7.46 -1.70
    GloGen 30.44 21.77 8.67 29.47 21.86 7.61 -1.07
    Table 5
    Performance Evaluation of Model pre-trained on MIMIC dataset. Avg SBP and Avg DBP
    represent the mean SBP MAE and mean DBP MAE, respectively, and Avg refers to the
    sum of these two. For GA SBP and GA DBP, the mean is first calculated for each BP
    group individually for both SBP and DPB, and then these values are averaged across
    the BP groups. GA refers to the sum of SBP GA and DBP GA. 2
    refers to the coefficient
    of determination. All units are mmHg.
    Pre-trained: MIMIC - Target: VitalDB
    Shot Method Avg Avg SBP Avg DBP GA GA SBP GA DBP 2
    5
    Scratch 25.63 15.61 10.03 32.74 20.51 12.23 -0.48
    LP 25.69 15.74 9.94 32.06 20.58 11.48 -0.40
    FT 25.00 15.37 9.63 31.38 20.14 11.24 -0.32
    GloGen 23.41 14.21 9.20 30.69 19.53 11.17 -0.04
    10
    Scratch 24.04 14.94 9.10 33.88 22.31 11.57 -0.14
    LP 26.11 16.20 9.91 31.47 20.41 11.06 -0.46
    FT 26.58 16.07 10.51 31.38 19.92 11.46 -0.61
    GloGen 23.78 14.49 9.29 29.48 18.76 10.71 -0.08
    4.4. The role of GloGen prompt
    Even when SBP and DBP values are similar, the shapes of PPG sig￾nals are not identical due to a variety of factors, such as measurement
    environments or individual patient characteristics (Fig. 4(a)). GloGen
    Prompt assists the pre-trained model in successfully performing BP
    estimation by making the shape of unseen PPG signals resemble those
    in the training set with similar SBP and DBP. In Fig. 4(a, b), we observe
    significant increases in the similarity between PPG signals from the
    training set and the test set that have comparable SBP and DBP. It not
    only occurs among PPG signals within the same BP groups but also
    across different BP groups, indicating the effective operation of GP in
    learning shared features in the training set.
    However, it is still evident that PPG signals within the same BP
    group, exhibit greater similarity. It demonstrates that Prompt Generator
    successfully utilizes BP group information when generating diverse IPs.
    Fig. 4(c) demonstrates how IP utilizes information about BP groups to
    transform the PPG signal, facilitating more robust BP estimation. As dis￾cussed in previous section, pre-trained models generally perform better
    in LG than in HG. GloGen encourages diversity among IPs through the
    regularization by VP. It ensures that IP adequately learns from BP group
    information. In Fig. 4(c), IPs are distinctly separated into clusters for
    HG and LG, indicating that the mechanism by which IP modifies the
    input signal varies between HG and LG. Moreover, within HG, IPs are
    differentiated based on BP groups, Hypo and Hyper2. Consequently,
    Prompt Generator generates IPs characterized by different features
    according to BP groups, resulting in GloGen’s ability to transform the
    PPG signal for robust BP estimation.
    4.5. Ablation studies
    Fig. 3 illustrates how the size and presence of GP and IP impact
    the model’s Average MAE and Group Average MAE performance. In
    the experiment where the model is pre-trained on the UCI dataset
    and fine-tuned with only 5-shot learning on the Sensors dataset, we
    Computers in Biology and Medicine 183 (2024) 109216
    6
    T. Kim et al.
    Fig. 4. (a) Comparison of shape and cosine similarity between two PPG signals that are present in the training set and the test set, respectively, each having similar SBP and
    DBP values. (b) Changes in shape and cosine similarity following the addition of the corresponding GloGen prompt to the PPG signals presented in (a). (c) UMAP visualization of
    the distribution for IPs generated by Prompt Generator for the test set, following transfer learning to the target task UCI dataset in 10-shot setting using GloGen, with the model
    pre-trained on Sensors dataset.
    Table 6
    Performance evaluation of GloGen on classification. All metrics are
    reported as percentages.
    Shot Method Acc Group Acc Precision Recall F1-Score
    5
    Scratch 41.7 29.3 32.2 29.3 29.5
    LP 23.3 37.1 26.9 37.1 17.0
    FT 31.3 36.1 40.7 36.1 21.7
    GloGen 43.5 43.0 57.7 43.0 28.2
    10
    Scratch 42.9 26.1 24.3 26.1 23.0
    LP 16.7 34.1 22.2 34.1 15.1
    FT 17.7 34.9 22.2 34.9 15.1
    GloGen 42.3 40.8 30.0 40.8 26.9
    observe performance changes on the Sensors dataset by adjusting the
    hyperparameters 𝜆 and 𝛾, which determine the sizes of GP and IP.
    We also analyze scenarios where GP or IP is absent (N/A). GloGen
    achieves the best Average MAE and Group Average performance when
    both GP and IP are appropriately utilized. It demonstrates that both
    instance-specific information (IP) and common information across the
    dataset (GP) are necessary for robust prediction. Table 4 demonstrates
    that the presence or absence of GP normalization, 𝑥
    𝐺𝑃
    , 𝑋𝐼𝑃 , and VP
    significantly influences GloGen’s performance. It also highlights that
    they each play a critical role in enhancing both the performance and
    robustness of GloGen.
    4.6. Pre-trained on a large-scale dataset (MIMIC)
    With the advancement of Electronic Health Records, extensive med￾ical data collection has become possible, leading to the creation of
    large-scale datasets like the MIMIC dataset. Transfer learning aims to
    effectively utilize the knowledge from pre-trained models on different
    target datasets, making few-shot transfer learning powerful by not
    requiring a large-scale target dataset. Therefore, it is important to
    verify whether a transfer learning algorithm shows valid performance
    improvements when applied to a large-scale pre-trained model. We
    apply GloGen to a model pre-trained on MIMIC to evaluate whether
    GloGen can effectively facilitate the knowledge of the MIMIC pre￾trained model, enhancing the model’s generalization performance and
    robustness regarding BP groups. The target dataset for this experiment
    is VitalDB [34].
    In medical data experiments, the performance of the model can
    significantly change depending on whether the same patients exist
    between the train and test sets, potentially leading to information
    leakage about the subjects. MIMIC and VitalDB are non-overlapping
    datasets regarding patient information, and the training and validation
    sets for VitalDB are also set to avoid overlapping with the test set
    Table 7
    Performance of GloGen and all baselines varies with the number of peaks in PPG signals.

of Peaks Scratch LP FT GloGen

4 41.5 ± 2.1 25.2 ± 4.3 26.6 ± 5.5 35.4 ± 4.7
5 35.2 ± 4.4 19.3 ± 5.2 28.7 ± 10.2 7.4 ± 6.3
8 99.6 ± 2.3 82.9 ± 2.0 97.4 ± 0.8 77.8 ± 15.3
subjects [34]. Table 5 shows the results. Except for the average DBP
MAE in the 10-shot setting, GloGen outperforms the baseline across
all other metrics. Notably, it significantly improves Group Average
performance, indicating that GloGen remains effective even for large￾scale pre-trained models and successfully learns a robust model for
blood pressure group changes.
4.7. Classification
GloGen excels not only in regression tasks for SBP and DBP but
also in classification tasks that directly predict BP groups. To assess
GloGen’s effectiveness across various metrics, we include average accu￾racy, group accuracy (the average of accuracy in each group), precision,
recall, and F1-score in Table 6. We evaluate BP classification per￾formance on the BCG dataset in the few-shot setting using a model
pre-trained on the Sensors dataset. GloGen achieves the best perfor￾mance in all metrics except for average accuracy in the 10-shot setting,
demonstrating that the transformation of PPG signals with GloGen is
versatile and applicable to different tasks.
4.8. Analysis of failure cases
We observe that the number of peaks in PPG signals significantly
impacts model performance. Fig. 5 illustrates a portion of the BCG data,
while Table 7 presents the MAE values calculated when the model,
pre-trained on the UCI dataset and fine-tuned with the 5-shot BCG
dataset, estimates the data shown in Fig. 5. Except for the Fine-tuned
(FT) model, the other models demonstrate the best performance on the
input PPG signal in Fig. 5 with 5 peaks. Fig. 5(c) depicts data with 8
peaks, where all baselines, including GloGen, exhibit very high MAE
values.
4.9. Discussion
Numerous studies have proposed novel model architectures to pre￾dict SBP and DBP solely from PPG signals [4,9,22,23,36]. [13] com￾pares the average MAE results of three representative architectures –
ResNet1D [23], MLPBP [22], and SpectroResNet [4] – across various
Computers in Biology and Medicine 183 (2024) 109216
7
T. Kim et al.
Fig. 5. PPG signals from the BCG dataset categorized by the number of peaks.
datasets. Among these, ResNet1D is selected as the pre-trained model
in our study due to its overall superior performance. However, as
illustrated in Tables 1 and 5, the prediction model lacks robustness
in our setting where the pre-trained dataset and the target dataset
differ. Traditional transfer learning methods, such as FT and LP, could
be employed to address this issue, but their robustness is inconsistent
depending on the combination of the pre-trained and target distribu￾tions (Fig. 2). GloGen, on the other hand, consistently demonstrates
superior performance and robustness, irrespective of the distributional
or dataset size differences between the train and target datasets.
The primary objective of transfer learning is to ensure that knowl￾edge obtained during pre-training can be effectively applied to target
datasets with different distributions. While recent BP estimation studies
using PPG signals have explored transfer learning, most have presented
limited results by confining the transfer learning to the same dataset.
For instance, [8,9] separate pre-train and target datasets by patient
ID within the MIMIC-III dataset, requiring additional demographic
information, which restricts the applicability of the framework to spe￾cific datasets. Similarly, [37] proposes a transfer learning framework
confined within the MIMIC-II dataset. [38] introduces a contrastive
learning-based pre-training framework using large-scale datasets, pair￾ing high-quality and low-quality PPG signals. While this framework
presents results for downstream tasks with target datasets substantially
different from the pre-trained dataset, it relies on large-scale data for
pre-training and lacks quantitative evaluations of model robustness
across different BP ranges. While previous studies have focused on full
fine-tuning for the target dataset, our experimental results in the few￾shot setting, however, show that GloGen significantly enhances model
robustness with only a small number of samples in target distribution.
Additionally, the prompt-based approach, which is not constrained by
specific model architectures, further highlights the potential impacts of
our studies.
5. Conclusion
In this paper, we propose GloGen, a novel framework for few-shot
transfer learning on BP estimation using PPG signal. We explicitly
present the performance differences across BP groups for representative
transfer learning methods and demonstrate the effectiveness of prompt
learning using GloGen. Thorough GP for the shared features and IP for
the instance-specific features, GloGen successfully manages the trade￾off between BP estimation performance on the target distribution and
robustness across different BP groups. We also introduce VP to en￾courage Prompt Generator to generate diverse IPs. As a result, Prompt
Generator effectively extracts BP group information corresponding to
each PPG signal and IP facilitates robust BP estimation by transforming
the PPG signals. GloGen is not only applicable to PPG signals but can
also be extended to all other signal data, which we aim to explore in
future work.
CRediT authorship contribution statement
Taero Kim: Writing – review & editing, Writing – original draft, Vi￾sualization, Validation, Software, Methodology, Formal analysis, Con￾ceptualization. Hyeonjeong Lee: Writing – review & editing, Project
administration, Methodology, Investigation. Minseong Kim: Writing –
review & editing, Visualization, Software. Kwang-Yong Kim: Writing –
review & editing, Funding acquisition, Conceptualization. Kyu Hyung
Kim: Writing – review & editing, Funding acquisition, Conceptualiza￾tion. Kyungwoo Song: Writing – review & editing, Writing – orig￾inal draft, Project administration, Investigation, Funding acquisition,
Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing finan￾cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Acknowledgments
This work was supported by Electronics and Telecommunications
Research Institute (ETRI) grant funded by the Korean government.
[24ZD1140, Regional Industry ICT Convergence Technology Advance￾ment and Support Project in Daegu-GyeongBuk (Medical)], and sup￾ported by the National Research Foundation of Korea (NRF) grant
funded by the Korea government (MSIT) (RS-2024-00457216), and the
grant (22183MFDS431) from Ministry of Food and Drug Safety in 2024,
and the Industrial Technology Alchemist Project Program (No. RS-
2024-00419010) funded By the Ministry of Trade, Industry & Energy
(MOTIE, Korea).
References
[1] X. Teng, Y. Zhang, Continuous and noninvasive estimation of arterial blood
pressure using a photoplethysmographic approach, in: Proceedings of the 25th
Annual International Conference of the IEEE Engineering in Medicine and Biology
Society (IEEE Cat. No. 03CH37439), Vol. 4, IEEE, 2003, pp. 3153–3156.
[2] W. Shi, C. Zhou, Y. Zhang, K. Li, X. Ren, H. Liu, X. Ye, Hybrid mod￾eling on reconstitution of continuous arterial blood pressure using finger
photoplethysmography, Biomed. Signal Process. Control 85 (2023) 104972.
[3] H. Samimi, H.R. Dajani, A PPG-based calibration-free cuffless blood pressure
estimation method using cardiovascular dynamics, Sensors 23 (8) (2023) 4145.
[4] G. Slapničar, N. Mlakar, M. Luštrek, Blood pressure estimation from photo￾plethysmogram using a spectro-temporal deep neural network, Sensors 19 (15)
(2019) 3420.
[5] M. Kim, H. Lee, K.-Y. Kim, K.-H. Kim, Deep learning model for blood pres￾sure estimation from PPG signal, in: 2022 IEEE International Conference on
Metrology for Extended Reality, Artificial Intelligence and Neural Engineering
(MetroXRAINE), IEEE, 2022, pp. 1–5.
[6] A. Tazarv, M. Levorato, A deep learning approach to predict blood pressure
from ppg signals, in: 2021 43rd Annual International Conference of the IEEE
Engineering in Medicine & Biology Society, EMBC, IEEE, 2021, pp. 5658–5662.
Computers in Biology and Medicine 183 (2024) 109216
8
T. Kim et al.
[7] G. Frederick, et al., PPG signals for hypertension diagnosis: A novel method using
deep learning models, 2023, arXiv preprint arXiv:2304.06952.
[8] Y. Zhang, X. Ren, X. Liang, X. Ye, C. Zhou, A refined blood pressure estimation
model based on single channel photoplethysmography, IEEE J. Biomed. Health
Inf. 26 (12) (2022) 5907–5917.
[9] Z. Liu, Y. Zhang, C. Zhou, Bigru-attention for continuous blood pressure trends
estimation through single channel PPG, Comput. Biol. Med. 168 (2024) 107795.
[10] C.-T. Yen, S.-N. Chang, C.-H. Liao, Deep learning algorithm evaluation of hy￾pertension classification in less photoplethysmography signals conditions, Meas.
Control 54 (3–4) (2021) 439–445.
[11] E.-S.A. El-Dahshan, M.M. Bassiouni, S.K. Khare, R.-S. Tan, U.R. Acharya, Ex￾HyptNet: An explainable diagnosis of hypertension using EfficientNet with PPG
signals, Expert Syst. Appl. (2023) 122388.
[12] E. Martinez-Ríos, L. Montesinos, M. Alfaro-Ponce, A machine learning approach
for hypertension detection based on photoplethysmography and clinical data,
Comput. Biol. Med. 145 (2022) 105479.
[13] S. González, W.-T. Hsieh, T.P.-C. Chen, A benchmark for machine-learning based
non-invasive blood pressure estimation using photoplethysmogram, Sci. Data 10
(1) (2023) 149.
[14] L. Breiman, Random forests, Mach. Learn. 45 (2001) 5–32.
[15] S. Maqsood, S. Xu, M. Springer, R. Mohawesh, A benchmark study of ma￾chine learning for analysis of signal feature extraction techniques for blood
pressure estimation using photoplethysmography (PPG), Ieee Access 9 (2021)
138817–138833.
[16] K. Duan, Z. Qian, M. Atef, G. Wang, A feature exploration methodology for learn￾ing based cuffless blood pressure measurement using photoplethysmography, in:
2016 38th Annual International Conference of the IEEE Engineering in Medicine
and Biology Society, EMBC, IEEE, 2016, pp. 6385–6388.
[17] G. Wang, M. Atef, Y. Lian, Towards a continuous non-invasive cuffless blood
pressure monitoring system using PPG: Systems and circuits review, IEEE Circuits
Syst. Magaz. 18 (3) (2018) 6–26.
[18] T. Treebupachatsakul, A. Boosamalee, S. Shinnakerdchoke, S. Pechprasarn, N.
Thongpance, Cuff-less blood pressure prediction from ecg and ppg signals using
Fourier transformation and amplitude randomization preprocessing for context
aggregation network training, Biosensors 12 (3) (2022) 159.
[19] Y. Yoon, J.H. Cho, G. Yoon, Non-constrained blood pressure monitoring using
ECG and PPG for personal healthcare, J. Med. Syst. 33 (2009) 261–266.
[20] B.L. Hill, N. Rakocz, Á. Rudas, J.N. Chiang, S. Wang, I. Hofer, M. Cannesson,
E. Halperin, Imputation of the continuous arterial line blood pressure waveform
from non-invasive measurements using deep learning, Sci. Rep. 11 (1) (2021)
15755.
[21] N. Ibtehaz, S. Mahmud, M.E. Chowdhury, A. Khandakar, M. Salman Khan, M.A.
Ayari, A.M. Tahir, M.S. Rahman, PPG2ABP: Translating photoplethysmogram
(PPG) signals to arterial blood pressure (ABP) waveforms, Bioengineering 9 (11)
(2022) 692.
[22] B. Huang, W. Chen, C.-L. Lin, C.-F. Juang, J. Wang, MLP-BP: A novel framework
for cuffless blood pressure measurement with PPG and ECG signals based on
MLP-mixer neural networks, Biomed. Signal Process. Control 73 (2022) 103404.
[23] F. Schrumpf, P. Frenzel, C. Aust, G. Osterhoff, M. Fuchs, Assessment of non￾invasive blood pressure prediction from ppg and rppg signals using deep learning,
Sensors 21 (18) (2021) 6022.
[24] B. Lester, R. Al-Rfou, N. Constant, The power of scale for parameter-efficient
prompt tuning, 2021, arXiv preprint arXiv:2104.08691.
[25] X.L. Li, P. Liang, Prefix-tuning: Optimizing continuous prompts for generation,
2021, arXiv preprint arXiv:2101.00190.
[26] P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, G. Neubig, Pre-train, prompt,
and predict: A systematic survey of prompting methods in natural language
processing, ACM Comput. Surv. 55 (9) (2023) 1–35.
[27] X. Liu, K. Ji, Y. Fu, W.L. Tam, Z. Du, Z. Yang, J. Tang, P-tuning v2: Prompt
tuning can be comparable to fine-tuning universally across scales and tasks, 2021,
arXiv preprint arXiv:2110.07602.
[28] M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, S.-N. Lim,
Visual prompt tuning, in: European Conference on Computer Vision, Springer,
2022, pp. 709–727.
[29] C. Oh, H. Hwang, H.-y. Lee, Y. Lim, G. Jung, J. Jung, H. Choi, K. Song, Blackvip:
Black-box visual prompting for robust transfer learning, in: Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp.
24224–24235.
[30] H. Bahng, A. Jahanian, S. Sankaranarayanan, P. Isola, Exploring visual prompts
for adapting large-scale models, 2022, arXiv preprint arXiv:2203.17274.
[31] K. Zhou, J. Yang, C.C. Loy, Z. Liu, Conditional prompt learning for vision￾language models, in: Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, 2022, pp. 16816–16825.
[32] K. Zhou, J. Yang, C.C. Loy, Z. Liu, Learning to prompt for vision-language
models, Int. J. Comput. Vis. 130 (9) (2022) 2337–2348.
[33] H. Xue, F.D. Salim, Promptcast: A new prompt-based learning paradigm for time
series forecasting, IEEE Trans. Knowl. Data Eng. (2023).
[34] W. Wang, P. Mohseni, K.L. Kilgore, L. Najafizadeh, PulseDB: A large, cleaned
dataset based on MIMIC-III and VitalDB for benchmarking cuff-less blood
pressure estimation methods, Front. Digit. Health 4 (2023) 1090854.
[35] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
2016, pp. 770–778.
[36] D. Dai, Z. Ji, H. Wang, Non-invasive continuous blood pressure estimation from
single-channel PPG based on a temporal convolutional network integrated with
an attention mechanism, Appl. Sci. 14 (14) (2024) 6061.
[37] H.M. Koparır, Ö. Arslan, Cuffless blood pressure estimation from photoplethys￾mography using deep convolutional neural network and transfer learning,
Biomed. Signal Process. Control 93 (2024) 106194.
[38] C. Ding, Z. Guo, Z. Chen, R.J. Lee, C. Rudin, X. Hu, SiamQuality: a ConvNet￾based foundation model for photoplethysmography signals, Physiol. Meas. 45 (8)
(2024) 085004.
Computers in Biology and Medicine 183 (2024) 109216
9

Réflexion

解释理论和应用

将LaTeX公式结合起来解释文章的方法,逐步详述经济学模型的定义和框架,并通过一个具体的例子进行说明。

思考,持续 几秒

Réponse

好的,下面我将结合论文中的LaTeX公式,详细说明《GloGen: PPG prompts for few-shot transfer learning in blood pressure estimation》一文的研究思路,并通过一个具体的例子来说明其方法。


1. 研究背景

血压(Blood Pressure, BP)监测是心血管健康管理的关键指标。传统的血压测量方法依赖于袖带或有创动脉血压测量(Arterial Blood Pressure, ABP),这些方法在实时、连续和无创监测方面存在诸多不便。光电容积描记图(Photoplethysmogram, PPG)是一种可通过简单传感器非侵入性获取的生理信号,已被证实与ABP存在一定相关性,因此利用PPG来预测血压有望实现便捷、实时的无创连续监测。

然而,现有基于PPG的BP预测模型在跨数据集或不同人群(如低血压和高血压高风险人群)时的泛化能力和鲁棒性仍然存在较大挑战。特别是在医疗数据领域,目标任务的数据常常稀缺(few-shot场景),这使得传统的迁移学习策略效用受限。

为了解决上述问题,论文提出了一种名为**GloGen(Global Prompt and Prompt Generator)**的新颖框架,旨在在极少数据条件下,通过prompt式迁移学习手段,将预训练模型对PPG信号的知识高效传递给目标任务,从而提高血压估计的准确性和对不同BP亚群体的鲁棒性。


2. GloGen框架的核心思想

GloGen的核心思想是利用Prompt Learning的方法,通过引入两种类型的提示(Prompts)——Global Prompt (GP)Instance-wise Prompt (IP),来调整输入的PPG信号,使预训练模型更好地适应目标任务的分布和特性,即使在few-shot的场景下也能获得良好的表现。

2.1 Global Prompt (GP)

GP是一组可训练参数,与任意输入PPG信号线性叠加(加法)以得到改造后的输入。公式表示为:

xGloGen,i=xi+λxGP+γxIP,i\mathbf{x}_{\text{GloGen}, i} = \mathbf{x}_i + \lambda \cdot \mathbf{x}_{\text{GP}} + \gamma \cdot \mathbf{x}_{\text{IP}, i}

其中:

  • xi\mathbf{x}_i 是第 ii 个PPG信号。
  • xGP\mathbf{x}_{\text{GP}} 是Global Prompt,λ\lambda 是控制GP影响的系数。
  • xIP,i\mathbf{x}_{\text{IP}, i} 是Instance-wise Prompt,γ\gamma 是控制IP影响的系数。

GP在所有样本上共享,用于捕捉目标数据集的全局特征或公共模式,帮助模型更好地适配总体分布。

2.2 Instance-wise Prompt (IP) & Prompt Generator

IP是为每个输入样本动态生成的特定“提示”,通过一个称为Prompt Generator的网络生成。具体表示为:

xIP,i=gϕh(xi)=gϕ(zi)\mathbf{x}_{\text{IP}, i} = g_\phi \circ h(\mathbf{x}_i) = g_\phi(\mathbf{z}_i)

其中:

  • h(xi)h(\mathbf{x}_i) 是特征提取器,将PPG信号映射到特征嵌入空间,得到 zi\mathbf{z}_i
  • gϕg_\phi 是Prompt Generator,由反卷积层、归一化层和激活函数(如ReLU)组成。

此外,论文引入了触发向量(Trigger Vector) Ztrigger\mathbf{Z}_{\text{trigger}},用于补充额外的可训练信息:

zi=[Zemb,Ztrigger]\mathbf{z}_i = [\mathbf{Z}_{\text{emb}}, \mathbf{Z}_{\text{trigger}}]

2.3 Variance Penalty (VP)

为了增强模型在不同BP子群体下的鲁棒性,GloGen设计了一个方差惩罚项(Variance Penalty, VP),鼓励生成的IP在不同BP组间具有多样性。VP的定义为:

VP=1D1Kj=1Dk=1K(μjIPxˉIP,j,k)2\text{VP} = \frac{1}{D} \cdot \frac{1}{K} \sum_{j=1}^{D} \sum_{k=1}^{K} (\mu_j^{\text{IP}} - \bar{\mathbf{x}}_{\text{IP}, j, k})^2

其中:

  • DD 是IP的维度(时间步数)。
  • KK 是BP组的数量。
  • μjIP=1Kk=1KxˉIP,j,k\mu_j^{\text{IP}} = \frac{1}{K} \sum_{k=1}^{K} \bar{\mathbf{x}}_{\text{IP}, j, k} 是每个时间步的IP组平均值。
  • xˉIP,j,k\bar{\mathbf{x}}_{\text{IP}, j, k} 是BP组 kk 在第 jj 个时间步的IP平均值。

为了避免训练不稳定,VP通过ReLU函数进行调整,最终的损失函数为:

L=MSE+αReLU(mVP)L = \text{MSE} + \alpha \cdot \text{ReLU}(m - \text{VP})

其中,α\alpha 是控制MSE和VP之间权重的超参数,mm 是控制ReLU函数阈值的超参数。


3. 具体例子说明GloGen的工作流程

为了更直观地理解GloGen的工作原理,我们通过一个具体的例子来说明其流程。

3.1 示例场景

假设我们有一个预训练的模型 fθpenfϕregf_{\theta_{\text{pen}}} \circ f_{\phi_{\text{reg}}},该模型已在大量PPG数据上训练好,用于血压(BP)的估计。现在,我们希望将该模型迁移到一个新的数据集(目标任务),但目标数据集只有极少量的标注样本(例如,每个BP组只有5个样本)。

3.2 GloGen的应用步骤

  1. 输入PPG信号:首先,我们获取目标任务的数据集中的一个PPG信号 xi\mathbf{x}_i

  2. 生成嵌入向量

    • 使用特征提取器 hhxi\mathbf{x}_i 映射到特征嵌入空间,得到 zi=h(xi)\mathbf{z}_i = h(\mathbf{x}_i)
    • 若使用触发向量,则 zi=[Zemb,Ztrigger]\mathbf{z}_i = [\mathbf{Z}_{\text{emb}}, \mathbf{Z}_{\text{trigger}}]
  3. 生成Instance-wise Prompt (IP)

    • 通过Prompt Generator gϕg_\phi 生成IP:xIP,i=gϕ(zi)\mathbf{x}_{\text{IP}, i} = g_\phi(\mathbf{z}_i)
  4. 叠加Prompts

    • 将GP和IP叠加到原始PPG信号上,得到GloGen改造后的输入: xGloGen,i=xi+λxGP+γxIP,i\mathbf{x}_{\text{GloGen}, i} = \mathbf{x}_i + \lambda \cdot \mathbf{x}_{\text{GP}} + \gamma \cdot \mathbf{x}_{\text{IP}, i}
  5. 血压估计

    • 将改造后的输入 xGloGen,i\mathbf{x}_{\text{GloGen}, i} 输入到预训练模型 ff,得到SBP和DBP的预测值: y^i=f(xGloGen,i)\hat{y}_i = f(\mathbf{x}_{\text{GloGen}, i})
  6. 损失计算与模型更新

    • 计算均方误差(MSE)损失: MSE=1Ni=1N(y^iyi)2\text{MSE} = \frac{1}{N} \sum_{i=1}^{N} (\hat{y}_i - y_i)^2 其中,yiy_i 是真实的SBP和DBP值。
    • 计算方差惩罚(VP): VP=1D1Kj=1Dk=1K(μjIPxˉIP,j,k)2\text{VP} = \frac{1}{D} \cdot \frac{1}{K} \sum_{j=1}^{D} \sum_{k=1}^{K} (\mu_j^{\text{IP}} - \bar{\mathbf{x}}_{\text{IP}, j, k})^2
    • 结合VP,最终的损失函数为: L=MSE+αReLU(mVP)L = \text{MSE} + \alpha \cdot \text{ReLU}(m - \text{VP})
    • 反向传播更新GP、Prompt Generator gϕg_\phi 和回归层 fϕregf_{\phi_{\text{reg}}},但冻结预训练模型的参数 fθpenf_{\theta_{\text{pen}}}

3.3 具体示例

假设目标任务数据集中有四个BP组:Hypo(低血压)、Normal(正常血压)、Prehyper(前高血压)和Hyper2(高血压)。每个BP组有5个样本。

  1. 选择一个样本:选择Hypo组中的一个PPG信号 x1\mathbf{x}_1

  2. 生成GP和IP

    • GP xGP\mathbf{x}_{\text{GP}} 是全局可训练参数,通过训练逐步学习到适合整个数据集的调整。
    • IP xIP,1\mathbf{x}_{\text{IP}, 1} 由Prompt Generator基于z1\mathbf{z}_1生成,反映该样本的特定特征。
  3. 叠加Prompts

    xGloGen,1=x1+λxGP+γxIP,1\mathbf{x}_{\text{GloGen}, 1} = \mathbf{x}_1 + \lambda \cdot \mathbf{x}_{\text{GP}} + \gamma \cdot \mathbf{x}_{\text{IP}, 1}

    其中,λ\lambdaγ\gamma 分别控制GP和IP的影响力。

  4. 输入预训练模型

    y^1=f(xGloGen,1)\hat{y}_1 = f(\mathbf{x}_{\text{GloGen}, 1})

    预测得到SBP和DBP值。

  5. 损失计算

    • 计算MSE损失: MSE=(y^1y1)2\text{MSE} = (\hat{y}_1 - y_1)^2
    • 计算VP,根据当前BP组和其他组的IP差异,确保不同BP组生成的IP具有足够的多样性。
  6. 更新参数

    • 通过最小化损失 L=MSE+αReLU(mVP)L = \text{MSE} + \alpha \cdot \text{ReLU}(m - \text{VP}),更新GP、Prompt Generator和回归层的参数。

通过这种方式,GloGen能够在极少数样本的情况下,通过全局和实例化的提示,提高模型在不同BP组上的预测准确性和鲁棒性。


4. 论文的实验与结果

4.1 实验设置

  • 数据集:使用三个基准数据集(BCG、Sensors和UCI)以及大规模数据集(MIMIC和VitalDB)进行评估。
  • BP组划分:将BP值按范围分为Hypo、Normal、Prehyper和Hyper2四组。
  • 训练设置:采用few-shot设定,如每个BP组仅有5或10个训练样本。
  • 模型架构:使用ResNet1D作为基础模型。
  • 评价指标
    • 平均绝对误差(MAE)
    • 分组平均误差(Group Average MAE)
    • 决定系数(R2R^2

4.2 关键实验结果

  • 总体性能:GloGen在大多数测试场景下均表现出领先的平均MAE和组平均MAE性能,尤其在5-shot这样的极端少样本学习条件下表现更加明显。
  • 高风险组:相较于传统迁移学习方法,GloGen有效降低了高风险组(Hypo和Hyper2)的预测误差。
  • 鲁棒性:在跨数据集迁移学习中,GloGen显著提升了模型对不同BP分组的鲁棒性。

4.3 消融实验

通过去除GP、IP或VP,验证各组件对性能的贡献。结果显示,没有任何一个组件(GP、IP或VP)都会导致性能下降,表明各组件在GloGen框架中都起着关键作用。


5. 总结

论文提出的GloGen框架,通过引入全局提示(GP)和实例化提示(IP)的双Prompt策略,以及方差惩罚(VP),有效提升了在few-shot迁移学习场景下PPG信号的血压估计性能和模型的鲁棒性。通过具体的实验验证,GloGen在不同数据集和不同BP组上均表现出优越的性能,特别是在高风险BP组的预测中显著优于传统方法。

GloGen不仅适用于PPG信号,还具有良好的通用性和可扩展性,可应用于其他生理信号或时间序列预测任务中,为在真实医疗场景中实现高效、精准、可推广的无创血压估计提供了有力的工具。

Question

我还是看不懂,完全看不懂,对于一个初学者来说如何理解呢,这个是基于什么论文的方法,最开始是借鉴什么方法

Réflexion

这次我正在帮助用户更好地理解论文" GloGen:用于血压估计的少量样本传输学习"的内容,包括其基本方法和启发。

思考,持续 几秒

Réponse

理解这篇论文《GloGen: PPG prompts for few-shot transfer learning in blood pressure estimation》可能会有些复杂,尤其是对于初学者来说。下面,我将用更简单的语言,结合具体例子,逐步解释论文的核心思想、方法及其来源,帮助你更好地理解这项研究。


1. 研究背景

1.1 血压监测的重要性

血压是衡量心血管健康的关键指标。高血压和低血压都可能导致严重的健康问题,因此,准确、实时地监测血压非常重要。传统的血压测量方法(如袖带式测量)虽然准确,但不够方便,无法实现连续、实时的监测。

1.2 PPG信号是什么?

光电容积描记图(Photoplethysmogram, PPG)是一种通过简单的传感器(如手环或智能手表)非侵入性地获取的生理信号。PPG信号反映了血液流动和血容量的变化,与血压有一定的相关性。因此,利用PPG信号来估计血压是一种有前景的方法,能够实现便捷、实时的无创血压监测。


2. 研究问题

虽然使用深度学习模型从PPG信号预测血压已经取得了一些进展,但在以下几个方面仍存在挑战:

  1. 泛化能力不足:现有模型在面对不同数据集或不同人群(如高血压或低血压患者)时,性能会明显下降。
  2. 数据稀缺:在医疗领域,获取大量标注数据往往困难,因此在“少样本”(few-shot)场景下,模型难以有效学习。
  3. 不同血压组的平衡:模型不仅需要在整体上表现良好,还需要在不同血压组(如低血压、高血压)上有一致且准确的预测。

3. 研究方法:GloGen框架

为了应对上述挑战,论文提出了**GloGen(Global Prompt and Prompt Generator)**框架。这个框架的核心思想是通过“提示学习”(Prompt Learning)的方法,调整输入的PPG信号,使得预训练模型能够更好地适应新的任务,尤其是在数据量有限的情况下。

3.1 什么是Prompt Learning?

Prompt Learning最初来源于自然语言处理(NLP)领域。简单来说,它是一种在不大幅度修改预训练模型的情况下,通过添加“提示”(prompts)来引导模型完成特定任务的方法。这类似于给模型提供一些额外的信息或指示,帮助它更好地理解和处理新任务。

3.2 GloGen的两种提示

GloGen框架引入了两种类型的提示:

  1. 全局提示(Global Prompt, GP)

    • 作用:捕捉所有PPG信号中的共有特征,帮助模型适应整体数据分布。
    • 实现:GP是一组可训练的参数,它与每个输入的PPG信号相加。这样,无论输入是什么,GP都会对其进行统一的调整。
  2. 实例化提示(Instance-wise Prompt, IP)

    • 作用:捕捉每个具体PPG信号的独特特征,帮助模型适应个体差异。
    • 实现:IP是根据每个输入样本动态生成的,通过一个叫做Prompt Generator的网络生成。这个生成器根据输入信号的特征,生成适合该信号的独特提示。

3.3 方差惩罚(Variance Penalty, VP)

为了确保不同血压组的IP具有足够的多样性,GloGen引入了方差惩罚(VP)。这个机制鼓励生成的IP在不同血压组之间有明显的差异,从而提高模型在各个组上的鲁棒性。

3.4 具体流程

以下是GloGen框架的具体步骤:

  1. 输入PPG信号:获取一个PPG信号样本 xi\mathbf{x}_i
  2. 生成嵌入向量:通过特征提取器将 xi\mathbf{x}_i 转换为一个特征向量 zi\mathbf{z}_i
  3. 生成IP:使用Prompt Generator gϕg_\phi 根据 zi\mathbf{z}_i 生成实例化提示 xIP,i\mathbf{x}_{\text{IP}, i}
  4. 叠加GP和IP:将全局提示和实例化提示与原始PPG信号相加,得到改造后的输入信号: xGloGen,i=xi+λxGP+γxIP,i\mathbf{x}_{\text{GloGen}, i} = \mathbf{x}_i + \lambda \cdot \mathbf{x}_{\text{GP}} + \gamma \cdot \mathbf{x}_{\text{IP}, i} 其中,λ\lambdaγ\gamma 是控制GP和IP影响力的系数。
  5. 血压估计:将改造后的输入 xGloGen,i\mathbf{x}_{\text{GloGen}, i} 输入到预训练模型 ff,得到预测的收缩压(SBP)和舒张压(DBP): y^i=f(xGloGen,i)\hat{y}_i = f(\mathbf{x}_{\text{GloGen}, i})
  6. 损失计算与模型更新:计算预测误差(如均方误差 MSE)和方差惩罚,更新GP和Prompt Generator的参数,但冻结预训练模型的参数。

4. 具体例子

为了更直观地理解GloGen的工作原理,下面通过一个具体的例子来说明:

4.1 示例场景

假设你有一个预训练的深度学习模型,它已经在大量PPG信号数据上训练好,用于预测血压。现在,你希望将这个模型应用到一个新的数据集,但新的数据集只有极少量的标注样本(例如,每个血压组只有5个样本)。

4.2 GloGen的应用步骤

  1. 选择一个样本:从新的数据集中选择一个低血压(Hypo)组的PPG信号 x1\mathbf{x}_1
  2. 生成嵌入向量:使用特征提取器将 x1\mathbf{x}_1 转换为特征向量 z1\mathbf{z}_1
  3. 生成IP:通过Prompt Generator gϕg_\phi 根据 z1\mathbf{z}_1 生成实例化提示 xIP,1\mathbf{x}_{\text{IP}, 1}
  4. 叠加GP和IP xGloGen,1=x1+λxGP+γxIP,1\mathbf{x}_{\text{GloGen}, 1} = \mathbf{x}_1 + \lambda \cdot \mathbf{x}_{\text{GP}} + \gamma \cdot \mathbf{x}_{\text{IP}, 1}
    • 这里,λ\lambdaγ\gamma 可以根据需要调整,以控制GP和IP的影响。
  5. 输入预训练模型:将改造后的PPG信号 xGloGen,1\mathbf{x}_{\text{GloGen}, 1} 输入到预训练模型 ff,得到预测的SBP和DBP值。
  6. 损失计算与模型更新
    • 计算预测值与真实值之间的误差(如MSE)。
    • 计算VP,确保不同血压组的IP有足够的差异。
    • 结合误差和VP,更新GP和Prompt Generator的参数。

通过这个过程,GloGen能够在只有少量样本的情况下,通过全局和实例化的提示,显著提升模型在不同血压组上的预测准确性和鲁棒性。


5. GloGen的方法来源与借鉴

GloGen的方法主要借鉴和扩展了以下几个领域的研究成果:

  1. Prompt Learning(提示学习)

    • 来源:最初在自然语言处理(NLP)领域,用于指导预训练语言模型完成特定任务。
    • 借鉴:GloGen将Prompt Learning的概念引入到生理信号处理领域,通过添加提示来调整PPG信号,使模型更好地适应新任务。
  2. Transfer Learning(迁移学习)

    • 来源:广泛应用于各种机器学习任务,通过将一个任务上学到的知识迁移到另一个相关任务上,以提升新任务的学习效果。
    • 借鉴:GloGen利用预训练模型的知识,通过少量目标任务的数据(few-shot),实现对新数据集的适应。
  3. Computer Vision(计算机视觉)中的Prompt Learning

    • 来源:在计算机视觉领域,Prompt Learning被用于图像识别任务,通过添加可训练的提示来调整图像输入,使预训练的视觉模型更好地适应新任务。
    • 借鉴:GloGen采用类似的方法,但应用于PPG时间序列信号,通过添加全局和实例化提示来调整输入信号。
  4. Variance Penalty(方差惩罚)

    • 来源:在机器学习中,正则化技术(如方差惩罚)用于控制模型的复杂度,防止过拟合。
    • 借鉴:GloGen引入方差惩罚,鼓励不同血压组生成具有多样性的IP,从而提升模型的鲁棒性。

6. 实验与结果

6.1 实验设置

  • 数据集:使用多个基准数据集(如BCG、Sensors、UCI)以及大规模数据集(如MIMIC、VitalDB)。
  • 血压组划分:将数据集中的样本按照血压值分为低血压(Hypo)、正常血压(Normal)、前高血压(Prehyper)、高血压(Hyper2)四组。
  • 模型架构:采用ResNet1D(一种适合处理时间序列数据的深度学习模型)。
  • 评估指标
    • 平均绝对误差(MAE):衡量预测值与真实值之间的平均绝对差。
    • 分组平均误差(Group Average MAE):衡量模型在不同血压组上的平均误差,反映模型的鲁棒性。
    • 决定系数(R2R^2):衡量预测值与真实值的相关性。

6.2 关键实验结果

  • 整体性能提升:GloGen在大多数测试场景下的MAE和Group Average MAE都优于传统方法(如从头训练、仅微调最后一层、全面微调预训练模型)。
  • 高风险组表现更好:对于高风险血压组(Hypo和Hyper2),GloGen显著降低了预测误差,表现出更高的鲁棒性。
  • 少样本条件下尤为有效:在只有5个样本的极端少样本条件下,GloGen仍然能够取得显著的性能提升。

6.3 消融实验

为了验证各组件(GP、IP、VP)的重要性,研究人员进行了消融实验,即分别去除GP、IP或VP,观察模型性能的变化。结果显示,去除任何一个组件都会导致性能下降,证明这些组件在GloGen框架中都是必不可少的。


7. 总结与意义

7.1 主要贡献

  1. 创新的Prompt Learning框架:将Prompt Learning的理念从NLP扩展到PPG信号处理,提出了GloGen框架,结合了全局提示和实例化提示。
  2. 提升少样本迁移学习能力:GloGen在数据量极少的情况下,依然能够有效提升血压估计的准确性和鲁棒性。
  3. 改善高风险人群的预测:通过多样化的实例化提示,GloGen显著提高了对低血压和高血压群体的预测准确性。
  4. 广泛的适用性:GloGen不仅适用于PPG信号,还具有良好的通用性,能够扩展到其他生理信号或时间序列预测任务中。

7.2 实际应用价值

GloGen框架为在医疗领域实现高效、精准、可推广的无创血压估计提供了有力的工具,特别是在数据稀缺和需要对不同人群进行准确监测的实际场景中,具有重要的应用潜力。


8. 如何进一步学习和理解

对于初学者来说,以下几个步骤可以帮助你更好地理解GloGen及其相关概念:

  1. 基础知识

    • 机器学习与深度学习:了解基本的机器学习和深度学习概念,如神经网络、训练与测试、损失函数等。
    • 时间序列分析:学习如何处理和分析时间序列数据,这是理解PPG信号处理的基础。
  2. 核心概念

    • 迁移学习:学习迁移学习的基本原理,理解如何将一个任务上学到的知识应用到另一个相关任务中。
    • Prompt Learning:了解Prompt Learning在NLP和计算机视觉中的应用,理解如何通过添加提示来引导预训练模型完成特定任务。
  3. 具体方法

    • GloGen框架:详细阅读GloGen框架的步骤,理解全局提示和实例化提示的作用及其生成方式。
    • 正则化技术:学习正则化方法,特别是方差惩罚,理解其在提升模型鲁棒性中的作用。
  4. 实践练习

    • 动手实验:尝试使用开源代码或自己实现GloGen框架,应用到简单的时间序列数据集上,亲自体验其工作流程。
    • 阅读相关论文:阅读GloGen引用的相关论文,深入理解其理论基础和方法来源。

通过以上步骤,你将能够逐步掌握GloGen框架的核心思想和实现方法,并能够应用到实际的血压估计或其他生理信号处理任务中。


希望以上解释能够帮助你更好地理解这篇论文的内容和方法。如果还有具体的问题或需要进一步的解释,请随时提问!

Partager cette Q&R