Photoplethysmography (PPG) is a key signal for wearable health monitoring and clinical decision support. Existing universal PPG representation methods are mostly trained on high‑quality, large‑scale hospital oximeter data, and they struggle with sparse, noisy real‑world wearable recordings, limiting their generalization to clinically relevant tasks. Surprisingly, even after full fine‑tuning, PPG foundation models often fail to consistently outperform generic time‑series models.
Empirical studies on several PPG foundation models reveal inherent sparsity and computational redundancy, indicating that these models have already learned to activate task‑specific sub‑network structures. To preserve this valuable prior knowledge, we introduce the Structured Pruning‑and‑Fine‑Tuning framework (SPAFT). SPAFT first prunes the model based on parameter importance and structural redundancy, retaining a compact sub‑network relevant to the target task. Then, adaptive fine‑tuning is performed within this reduced parameter space, guiding optimization toward more pertinent directions while maintaining the learned priors.
Extensive experiments across four PPG foundation models, five benchmark datasets, and ten downstream tasks show that SPAFT achieves win rates of 82.8% on regression and 87.5% on classification tasks. The prune‑then‑fine‑tune paradigm often pushes the models to state‑of‑the‑art performance, offering a promising route for real‑world clinical deployment of PPG models and the development of personalized PPG foundations.
Review