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

SPOTR, a new self-supervised learning framework, significantly advances physiological signal processing by using a single-token bottleneck to compress and reconstruct EEG, ECG, PPG, and iEEG signals. The model demonstrates substantial performance improvements across 20 datasets while reducing computational requirements by 78% in latency and 52% in GPU memory compared to existing foundation models.

SPOTR addresses a critical gap in medical AI where current self-supervised learning methods struggle with physiological signal diversity and fail to preserve clinically meaningful patterns. The framework's innovation lies in its compress-reconstruct architecture, which forces the model to learn robust representations by condensing entire waveforms into single-token representations before reconstruction. This design principle eliminates the computational burden of processing flattened spatiotemporal sequences that plague transformer-based approaches.

The broader context reveals a sector-wide challenge: existing SSL methods exploit temporal and cross-channel redundancy as shortcuts, undermining clinical utility. SPOTR's multi-modal pretraining across 20 heterogeneous datasets—spanning four physiological signal types—demonstrates generalization capability missing from single-modality approaches. The dramatic performance gains under linear probing (18.49% AUC improvement for EEG) prove the method learns generalizable features rather than dataset-specific artifacts, a critical requirement for real-world medical deployment.

For medical AI developers and healthcare institutions, SPOTR's efficiency gains carry substantial implications. Reduced latency enables real-time clinical monitoring applications, while lower GPU memory requirements democratize deployment across resource-constrained hospital environments. The 52% GPU memory reduction particularly enables edge computing scenarios critical for portable medical devices. The open-source release creates an immediate foundation for downstream applications in cardiac monitoring, neurological assessment, and emergency diagnostics.

Future development hinges on clinical validation and integration into existing medical workflows. The framework's performance improvements must translate to diagnostic accuracy in prospective clinical trials. Adoption barriers will center on regulatory approval and integration with legacy hospital systems rather than technical feasibility.

• →SPOTR's single-token bottleneck design compresses physiological signals while preserving clinically meaningful patterns better than existing SSL methods
• →Performance improvements average 18.49% AUC gain for EEG and 21.71% for iEEG under linear probing evaluation
• →Computational efficiency reduces latency by 78% and GPU memory by 52% compared to general-purpose time-series foundation models
• →Multi-modal pretraining across 20 datasets spanning EEG, iEEG, ECG, and PPG demonstrates strong cross-signal generalization
• →Open-source release enables rapid adoption for medical device manufacturers and healthcare AI developers