GPT-PPG: A GPT-based Foundation Model for Photoplethysmography Signals

TL;DR

This work presents GPT-PPG, a GPT-based foundation model adapted for continuous PPG signals, pre-trained on a large corpus of over $2\times10^8$ thirty-second PPG samples. It introduces a mixed-objective fine-tuning framework combining task-specific objectives with a generative signal-modeling loss, enabling effective downstream tasks such as atrial fibrillation detection and physiological parameter estimation while supporting signal denoising. The model demonstrates strong performance on AF detection, HR and RR estimation, and BP prediction, with notable gains when distributions align and enhanced personalization improves generalization; it also reveals limited out-of-distribution generalization and reveals practical PEFT and bidirectional extensions to improve efficiency and robustness. The paper highlights the practical implications of applying foundation models to physiological data, outlines design choices like logit-Laplace loss and patch-based inputs, and discusses limitations and future directions, including broader pre-training data and patch-free sequence modeling for better generalization and deployment.

Abstract

This study introduces a novel application of a Generative Pre-trained Transformer (GPT) model tailored for photoplethysmography (PPG) signals, serving as a foundation model for various downstream tasks. Adapting the standard GPT architecture to suit the continuous characteristics of PPG signals, our approach demonstrates promising results. Our models are pre-trained on our extensive dataset that contains more than 200 million 30s PPG samples. We explored different supervised fine-tuning techniques to adapt our model to downstream tasks, resulting in performance comparable to or surpassing current state-of-the-art (SOTA) methods in tasks like atrial fibrillation detection. A standout feature of our GPT model is its inherent capability to perform generative tasks such as signal denoising effectively, without the need for further fine-tuning. This success is attributed to the generative nature of the GPT framework.

Figures (6)

• Figure 1: Mixed-objective Fine-tuning Framework
• Figure 2: Average Signal Modeling Loss. The datasets that are boxed in red are all heart rate estimation tasks. It is clear that on those datasets, our model is less competitive against the baselines compared to our model's performance on other datasets.
• Figure 3: NLL v.s. Performance on AF Detection and HR Estimation. Performance of AF detection is measured in F1 score (higher the better) and that of HR estimation is measured in MAE (lower the better). In both datasets, we observe that the performance generally gets worse as NLL increases.
• Figure 4: Noisy Signal Reconstruction Demonstrations
• Figure 5: Reconstruction Error v.s. Mask Probability

Theorems & Definitions (1)

• proof