Position: Evaluation of ECG Representations Must Be Fixed
A new position paper published on arXiv argues that the current methods for evaluating ECG representation learning need significant improvement. The paper highlights that the field has overly focused on a few specific benchmarks and label types, neglecting broader clinical information encoded in ECGs. Researchers propose expanding evaluations to include structural heart disease and patient-level forecasting, alongside best practices for handling imbalanced datasets, which could alter current conclusions about model performance. AI
IMPACT Proposes new evaluation standards for AI in healthcare, potentially improving the reliability and clinical relevance of diagnostic models.