Robust Epilepsy Diagnosis in Neonates Via Spatio-Temporal Graph Learning and Balanced Test-Time Training

Timely diagnosis of neonatal epilepsy is critical for mitigating long-term neurodevelopmental deficits. However, while neonates exhibiting extensive seizure burden are readily identified, accurately diagnosing epileptic infants with sparse or no overt seizure onset during monitoring remains a formidable challenge using electroencephalogram (EEG) signals alone. Current automated systems predominantly target ictal-specific seizure detection, limiting their clinical utility for risk stratification during seizure-free intervals (interictal), particularly given the pronounced EEG variability across neonates. This work presents a cross-subject deep learning framework that robustly distinguishes epileptic from "healthy" neonates by integrating dynamic graph topology with self-supervised adaptation. Our architecture models multichannel EEG as spatio-temporal graphs to capture functional connectivity, while a balanced test-time training mechanism dynamically adapts the model to unseen patients via masked signal reconstruction. This effectively mitigates the domain shifts caused by inter-subject variability without requiring patient-specific labels. Validated on a clinically curated cohort of 140 neonates (70 healthy, 70 epileptic), our approach achieves a segment-level accuracy of 83$\%$ and an area under the curve of 0.90. For the primary subject-level endpoint, accuracy increases from 0.79 without adaptation to 0.88 with balanced TTT, using aggregation thresholds selected on inner validation subjects and fixed before evaluation on each held-out fold. The cohort includes anoxic-ischemic encephalopathy, benign neonatal convulsions, and cerebrovascular accident. Analyses restricted to non-ictal windows also support discrimination between the two groups, although all neonates in the epilepsy group experienced seizures elsewhere in their recordings.