TPVNet: A domain-aware graph-based framework for reliable multivariate physiological time series classification in healthcare

Multivariate physiological time series classification is essential for healthcare decision support within the Internet of Medical Things (IoMT). However, existing methods often struggle with high noise, non-stationarity, and privacy concerns inherent in medical signals. This study proposes TPVNet, a novel domain-aware framework, to enhance classification accuracy, stability, and privacy protection in IoMT applications.

We propose TPVNet, a domain-aware graph-based framework specifically designed for physiological MTSC tasks. TPVNet integrates two key innovations: a Temporal-enhanced limited Penetrable Visibility Graph (TPVG) for converting raw signals into irreversible graph representations with enriched local and global temporal features. Graph Isomorphism Network (GIN) is used for feature learning followed by a channel-wise voting strategy that aligns with clinical diagnostic workflows to improve decision robustness. Experiments are conducted on seven public physiological datasets, comparing TPVNet against eight state-of-the-art baselines. Performance is evaluated using accuracy, recall, precision, F1-score and standard deviation over 10 replicate tests.

TPVNet demonstrates robust performance, achieving the highest F1-score on 6 out of 7 datasets. In terms of clinical utility, it significantly outperforms state-of-the-art baselines in data-scarce scenarios; notably, on the Atrial Fibrillation (AF) dataset, it boosts accuracy by 22.0%. Moreover, ablation studies confirm a cumulative accuracy gain of 12.7% over the graph baseline, validating the synergistic effectiveness of the proposed TPVG representation and voting mechanism. Furthermore, TPVNet exhibits superior stability, consistently yielding lower standard deviations compared to deep learning baselines.

TPVNet provides a privacy-aware, accurate, and stable solution for multivariate physiological time series classification. By integrating domain-inspired graph construction and decision fusion, it offers a clinically aligned framework suitable for real-world IoMT applications, bridging the gap between algorithmic design and healthcare needs.