Time is continuous, why chunk it? Token-Free State-Space Models for ECG

Sukwon Kim, William Han, Ding Zhao

Keywords: Electrocardiogram, Time Series data, State-Space Models, Foundation Models, Representation Learning, Token-Free Modeling

Abstract:

Recent successes in foundation models for electrocardiograms (ECGs) are heavily constrained by tokenization and fixed-length input paradigms. These models rely on destructive heuristics like zero-padding or rigid tokenization, rendering them fragile and computationally restricted when applied to highly variable real-world clinical data. To challenge these paradigms, we investigate State-Space Models (SSMs) as a foundational architecture, demonstrating their capability to natively process raw signals without artificial chunking.
Using a simple SSM-based encoder as a proof of concept, experiments show it achieves classification results competitive with state-of-the-art baselines on standard fixed-length benchmarks. Crucially, further evaluations on truncated and long signals reveal significantly less performance degradation, confirming the inherent robustness of SSMs architecture over fixed-length approaches.

Submission Number: 123