SafeECGMatch: Semi-Supervised ECG Classification with OOD Detection
A team of researchers has introduced SafeECGMatch, a framework designed for safe semi-supervised learning that is calibration-aware, specifically for single-label ECG classification amidst label distribution discrepancies. This approach tackles the challenge of mislabeling out-of-distribution anomalies found in unlabeled clinical data, which traditional SSL struggles with. Utilizing a dual-branch architecture, SafeECGMatch extracts time-frequency latent representations through ECG-specific augmentations. It aligns confidence with empirical accuracy dynamically by employing adaptive label smoothing and temperature scaling, effectively calibrating both the multiclass classifier and the OOD detector. The goal of this framework is to lower annotation expenses while ensuring dependable predictions in open-set situations.
Key facts
- SafeECGMatch is a calibration-aware safe SSL framework for ECG classification.
- It addresses label distribution mismatch with OOD anomalies in unlabeled data.
- Uses a dual-branch architecture for time-frequency latent representations.
- Employs adaptive label smoothing and temperature scaling for calibration.
- Calibrates both the multiclass classifier and the OOD detector.
- Aims to reduce annotation costs in clinical settings.
- Standard SSL forces incorrect pseudo-labels onto unseen classes.
- Framework is designed for single-label ECG classification.
Entities
Institutions
- arXiv