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

• Published in: Computers in biology and medicine 2024-12, Vol.183, p.109216, Article 109216
• Main Authors: Kim, Taero, Lee, Hyeonjeong, Kim, Minseong, Kim, Kwang-Yong, Kim, Kyu Hyung, Song, Kyungwoo
• Format: Article
• Language: English
• Subjects: Blood pressure; Blood Pressure - physiology; Blood Pressure Determination - methods; Blood pressure estimation; Female; Few-shot learning; Humans; Hypertension; Hypotension; Machine Learning; Male; Photoplethysmogram signal; Photoplethysmography - methods; Prompt learning; Real time; Risk groups; Robustness (mathematics); Signal Processing, Computer-Assisted; Telemedicine; Transfer learning
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2024.109216

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. [Display omitted] •We propose GloGen, a prompt learning method for blood pressure estimation using PPG.•GloGen guarantees both generalization and robustness across blood pressure groups.•Global Prompt in GloGen learns the shared features in the training set.•Prompt Generator in GloGen generates the personalized prompt for each PPG signal.•GloGen shows its efficacy in the few-shot transfer learning setting.