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

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Data Science

Background:

• Multivariate physiological time series classification is crucial for healthcare decision support in the Internet of Medical Things (IoMT).
• Existing methods face challenges with noisy, non-stationary medical signals and privacy concerns.
• There is a need for robust and privacy-preserving classification frameworks for IoMT.

Purpose of the Study:

• To introduce TPVNet, a novel domain-aware graph-based framework for multivariate physiological time series classification.
• To enhance classification accuracy, stability, and privacy protection in IoMT applications.
• To address limitations of existing methods in handling noisy and non-stationary medical data.

Main Methods:

• TPVNet utilizes a Temporal-enhanced limited Penetrable Visibility Graph (TPVG) to convert time series into irreversible graph representations with rich temporal features.
• Graph Isomorphism Network (GIN) is employed for feature learning.
• A channel-wise voting strategy, aligned with clinical workflows, enhances decision robustness.

Main Results:

• TPVNet achieved the highest F1-score on 6 out of 7 public physiological datasets.
• It significantly outperformed baselines in data-scarce scenarios, boosting Atrial Fibrillation (AF) classification accuracy by 22.0%.
• Ablation studies confirmed a 12.7% cumulative accuracy gain, demonstrating the effectiveness of the TPVG and voting mechanism. TPVNet also showed superior stability with lower standard deviations.

Conclusions:

• TPVNet offers a privacy-aware, accurate, and stable solution for multivariate physiological time series classification.
• The framework integrates domain-inspired graph construction and clinical decision fusion for real-world IoMT applications.
• TPVNet bridges the gap between advanced algorithmic design and practical healthcare needs.