SL-S4Wave: Self-Supervised Learning of Physiological Waveforms with Structured State Space Models

Analyzing long-sequence medical time series, such as ECG and EEG, is difficult due to high sampling rates, signal noise, and the scarcity of labeled data. Existing self-supervised learning (SSL) methods often rely on convolutional or recurrent architectures that struggle to capture long-range temporal dependencies or fail to handle the complexity of multichannel physiological data. This paper addresses these challenges by introducing a framework that combines contrastive learning with a specialized encoder based on structured state space models (S4).

The authors propose SL-S4Wave, which features a custom encoder called S4Wave. This encoder adapts structured state space models by incorporating multiscale global convolution kernels, residual connections, and cross-channel modeling. This architecture allows the model to extract both fine-grained local patterns and long-range temporal dynamics. The framework is trained using a contrastive objective that includes a noise-resilient loss—treating original and filtered signals as positive pairs—and a context-consistency loss that encourages temporal coherence between adjacent segments of the same physiological record.

SL-S4Wave demonstrates significant advantages over state-of-the-art baselines in several key areas:

• Performance: It consistently achieves superior accuracy in challenging arrhythmia detection tasks.
• Label Efficiency: The model maintains high performance even when provided with significantly fewer labeled examples, reducing the burden of manual annotation.
• Long-Sequence Modeling: Unlike CNN-based methods that degrade as sequence length increases, SL-S4Wave remains robust on long waveform segments, effectively capturing complex temporal dynamics.