HRVConformer: Neonatal Hypoxic-Ischemic Encephalopathy Classification from the Heart Rate signals

Researchers introduce HRVConformer, a deep learning model combining convolutional and Transformer architectures to classify neonatal hypoxic-ischemic encephalopathy (HIE) from heart rate signals. The model achieves 83.23% AUC and 74.56% accuracy, outperforming traditional baselines by automating HIE detection without requiring handcrafted features.

HRVConformer represents a meaningful advancement in medical AI by demonstrating how hybrid neural architectures can extract clinically relevant patterns from physiological signals. The model processes raw instantaneous heart rate data end-to-end, leveraging convolutional layers for local feature extraction and Transformer attention mechanisms for capturing long-range temporal dependencies—a combination that outperforms single-architecture approaches like ResNet50 and standard Transformers. This hybrid design proves particularly effective for HIE classification, a critical neonatal condition where early detection significantly impacts patient outcomes.

The research methodology reflects modern machine learning practices by incorporating both expert-annotated data (259 epochs) and weakly labelled datasets (1,573 total epochs), alongside a substantial 314-hour validation set. The use of an improved Pan-Tompkins algorithm for ECG signal preprocessing demonstrates attention to data quality, addressing a common bottleneck in medical AI. These practices increase model robustness and reproducibility, supported by the public code release.

For the broader medical AI field, this work validates that physiological signals contain sufficient discriminative information for automated disease classification when processed through appropriately designed architectures. The 74.56% accuracy on held-out test data suggests practical deployment readiness, though clinical validation remains necessary. This approach could extend to detecting other neonatal conditions or complications from continuous heart rate monitoring, potentially reducing clinician workload in intensive care settings.

Future development should focus on prospective clinical validation, integration with existing monitoring infrastructure, and investigation of model explainability to ensure clinical adoption and regulatory approval.

• →HRVConformer combines convolutional and Transformer layers to classify neonatal HIE from heart rate signals with 83.23% AUC, exceeding individual architecture baselines.
• →The hybrid architecture's strength lies in balancing local feature extraction via convolutions with global context modelling via attention mechanisms.
• →Training on 1,573 epochs including both expert-annotated and weakly labelled data demonstrates effective use of mixed-quality datasets in medical AI.
• →Improved Pan-Tompkins preprocessing enhanced signal quality and data availability, highlighting the importance of signal extraction methods in physiological AI.
• →Public code release enables reproducibility and supports broader adoption of the approach for similar physiological signal classification tasks.