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

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

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

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