Unlocking the Specialization Capability of PPG Foundation Models for Wearable Signals through Structured Pruning and Adaptive Fine-Tuning (SPAFT)

Anonymous (ICLR 2027 double-blind submission)

Summary

Finds that even fully fine-tuned PPG FMs do not consistently beat general time-series models, and that pretrained PPG FMs show inherent sparsity / redundancy. SPAFT constrains adaptive fine-tuning to a compact parameter subspace chosen by parameter importance and structural redundancy (prune-then-finetune). Across four PPG FMs, five benchmark datasets and ten downstream tasks: win rates 82.8% (regression) and 87.5% (classification).