SM: Bridging the Robustness Gap in Clinical Time Series Analysis via Hierarchical Stability Optimization

Songning Lai, Haoxuan Xu, Zhishan Tao, Wenshuo Chen, Ninghui Feng, Jianheng Tang, Haoang Li, Yutao Yue

Keywords: Medical Time Series, Robustness, EGG, ECG

TL;DR: SM boosts clinical reliability for medical time series by combining a clinician-inspired attention module with a multifaceted stability optimization strategy, enhancing robustness to real-world perturbations.

Abstract: Deep learning models for medical time series analysis exhibit a critical reliability gap: high accuracy on curated data does not translate to robustness against real-world noise and device variability. We argue this gap stems from inadequate modeling of hierarchical physiology and training paradigms that neglect clinical stability. We introduce $\textbf{SM}$ ($\textbf{S}$tability $\textbf{M}$edical time series classifier), a framework that bridges this gap by synergistically co-designing a novel, physiologically-inspired architecture with a multifaceted stability optimization strategy. Our $\textbf{S}$tability-aware $\textbf{H}$ierarchical $\textbf{S}$patial $\textbf{M}$odulation (SHSM) module mimics clinical reasoning by selectively attending to biomarkers while preserving global waveform morphology. Complementing this, our training objective enforces robust accuracy, output consistency, and knowledge preservation without sacrificing clean-data performance. Extensive evaluations on four medical time series datasets against 11 baselines demonstrate that SM achieves state-of-the-art performance while significantly improving robustness. By unifying architecture and training around the principle of stability, SM provides a systematic framework for building clinically reliable medical AI.

Primary Area: learning on time series and dynamical systems

Submission Number: 19116