Label-free Anomaly Detection in Cardiovascular Signals with Persistence-informed Multi-instance Learning

Area of Science:

• Cardiovascular Physiology
• Biomedical Signal Processing
• Machine Learning

Background:

• Detecting adverse physiological events in cardiovascular rhythms is crucial but challenged by "black-box" models and high annotation costs.
• Existing methods lack interpretability, hindering clinical trust and collaborative adaptation.

Purpose of the Study:

• To develop a label-free, interpretable model for enhanced cardiovascular event detection.
• To improve clinical trust and enable interactive collaboration in cardiovascular monitoring.

Main Methods:

• Proposed the Persistence-Informed Score-based Multi Instance Learning (PISMIL) framework.
• Utilized persistent homology for topological dynamics analysis and motif stability extraction.
• Employed confidence-weighted pseudo-labeling and an attention-based Multi Instance Learning (MIL) detector.

Main Results:

• PISMIL establishes a new benchmark in anomaly detection and motif localization across diverse ECG and PPG datasets.
• Demonstrated superior performance over established baselines on multiple evaluation metrics.
• Ablation studies confirmed the benefits of topological guidance and MIL for robustness and stability.

Conclusions:

• PISMIL offers a label-free, interpretable solution for cardiovascular event detection, addressing annotation burden and scalability.
• The framework generates anomaly maps and identifies discriminative motifs, facilitating clinical workflows.
• PISMIL has the potential for seamless integration into ubiquitous cardiovascular monitoring systems.