SPOTR: Spatio-temporal Pooling One-Token Reconstruction for Universal Physiological Signal Self-supervised Learning

SPOTR (Spatio-temporal Pooling One-Token Reconstruction) is a novel self-supervised learning framework designed for processing physiological signals like EEG, ECG, and PPG. It introduces a compress-reconstruct pretraining approach that reduces each waveform to a single-token representation, significantly lowering computational and memory costs while enhancing performance; SPOTR achieved improvements in average AUC of 18.49%, 21.71%, 17.86%, and 4.64% across four signal types under linear probing compared to strong baselines. This framework is particularly relevant for practitioners as it addresses the limitations of existing SSL methods in heterogeneous datasets and optimizes resource usage, making it suitable for real-world medical applications.