SegPPMTS: Unsupervised Segmentation for Pseudo-periodic Medical Time Series

Accurately segmenting medical time series into physiologically meaningful cycles is a critical preprocessing step to downstream tasks, such as classification. Unlike strictly periodic signals, medical time series, such as electrocardiograms and respiratory signals, exhibit pseudo-periodicity, characterised by variations in cycle length, amplitude, and morphology. The inherent variability of such medical signals poses challenges for existing segmentation approaches. To this end, we propose SegPPMTS, an unsupervised segmentation algorithm specifically designed for pseudo-periodic medical time series. SegPPMTS follows a three-step process: frequency-adaptive clustering to mitigate spectral leakage, periodicity hints validation to refine periodicity estimates, and dynamic time series segmentation to accommodate variations. It dynamically adjusts to pseudo-periodicity, enabling robust and adaptive segmentation across diverse signal types. Extensive experiments on four public medical datasets demonstrate that SegPPMTS outperforms existing segmentation methods across a comprehensive segmentation evaluation framework. Additionally, classification experiments validate our hypothesis that improved segmentation enhances classification performance, highlighting the importance of accurate pseudo-cycle segmentation of pseudo-periodic medical time series for downstream tasks. For clinical applications, our approach provides a promising solution for segmenting medical time series in scenarios where annotated segmentation ground truth is limited.