Functional Assessment of Neonatal Hypoxic–Ischemic Encephalopathy Using Long-Duration EEG and Interpretable Deep Learning Models

Neonatal hypoxic–ischemic encephalopathy (HIE) remains a critical neurological emergency resulting from perinatal asphyxia, often leading to lifelong neurodevelopmental disabilities or mortality. The accurate and timely grading of HIE severity is paramount for initiating therapeutic interventions such as therapeutic hypothermia. This work proposes a diagnostic framework that uses long-duration electroencephalogram (EEG) recordings through a hierarchical classification strategy and advanced sequence modeling. A Hybrid Mamba-inspired architecture was developed to effectively capture long-range temporal dependencies in multi-channel neonatal EEG while maintaining computational efficiency. In order to enhance clinical consistency and initialize the models appropriately, a Self-Supervised Learning step based on Masked Signal Modeling is implemented with a mask ratio of 30%. The model structure takes into consideration clinically verified biomarkers, including the suppression ratio, Delta–Alpha Ratio, Spectral Edge Frequency, and Rhythmicity Index, extracted from signals at a microvolt level prior to normalization for physiological interpretability purposes. These features are combined with waveforms using feature gating. In an experiment conducted on a dataset of 169 records using 5-fold subject-wise cross-validation, the designed Hybrid Mamba-based model achieves significant stability and generalizability, achieving an accuracy score of 90%, with an average accuracy of 88.45% ± 6.8% per hierarchical level.