Fetalet: interpretable fetal heart rate anomaly detection via shapelet learning

Abstract Fetal heart rate (FHR) monitoring is essential for detecting fetal distress, yet current deep learning approaches suffer from limited discriminative accuracy and poor interpretability, hindering clinical adoption. To address these challenges, we propose Fetalet, an interpretable framework for FHR anomaly detection with three key innovations: (1) shapelet-driven methodology capturing discriminative temporal patterns through learnable shapelets, (2) linear-time computational complexity via convolutional parameterization, and (3) clinically verifiable decision rationale through prototype-based metric learning. Our shapelet visualization approach enables direct clinical validation, with 92% alignment rate between automatically detected patterns and clinically recognized waveforms. Evaluated on both public (NIFEA) and private (FHRA) datasets, Fetalet achieves state-of-the-art performance with 0.98 F1-score and 0.96 accuracy while demonstrating robustness to clinical noise conditions, maintaining less than 3% performance degradation under Gaussian noise levels up to 0.2. Notably, on the FHRA dataset, Fetalet demonstrates a 19.7% F1-score improvement over Anomaly Transformer with significantly faster training time (3.6s versus 100 s). The framework’s interpretability features enable clinicians to validate model decisions through shapelet matching with established clinical patterns, enhancing trust in AI-assisted monitoring systems. This work bridges the gap between high-performance deep learning and clinically actionable interpretability, offering a promising solution for real-world fetal monitoring applications where both accuracy and transparency are critical for clinical decision-making.