Pseudo Anomalies and Hard Sample Mining for Ventricular Arrhythmia Anomaly Detection.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Ventricular arrhythmias (VA) are critical cardiac conditions requiring accurate detection.
• Conventional supervised methods struggle to differentiate VA from noise due to waveform similarities.
• Lack of large, annotated abnormal ECG datasets hinders supervised learning.

Purpose of the Study:

• To develop a novel anomaly detection framework (PHVA) for VA detection.
• To overcome the limitations of supervised learning by utilizing only normal ECG data.
• To improve the discrimination between VA and noise artifacts.

Main Methods:

• Implemented a one-class anomaly detection approach using only normal ECGs for training.
• Developed self-supervised modules for pseudo-anomaly generation and hard-sample mining.
• Utilized a physiology-aware synthetic ECG generation method and a time-frequency hypersphere model.
• Employed triplet-loss-based hard-sample mining for enhanced discriminative power.

Main Results:

• PHVA achieved superior overall performance compared to state-of-the-art anomaly detection methods.
• Outperformed baselines by up to 8.7% in Area Under the ROC Curve (AUC).
• Demonstrated effective discrimination between VA and noise artifacts.

Conclusions:

• The proposed PHVA framework offers a robust and effective solution for VA detection.
• One-class anomaly detection with pseudo-data generation is a viable alternative to supervised methods.
• PHVA shows significant potential for clinical application in arrhythmia monitoring.