Predicting long-term adverse outcomes after neonatal intensive care

Neonates requiring intensive care are at increased risk for long-term neuropsychiatric disorders. However, clinical adoption of risk prediction models remains limited when their performance lacks adequate interpretability for informed clinical decision-making. Here, we investigated whether longitudinal neonatal electronic health record (EHR) data from the first 90 days of life can support clinically meaningful interpretation of long-term risk signals for major neuropsychiatric diagnoses by age seven. In a retrospective register-based cohort of 17,655 at-risk children from an academic medical center, of whom 8.0\% (1,420) received a major neuropsychiatric diagnosis during follow-up, we applied a time-aware transformer model (Self-supervised Transformer for Time-Series; STraTS) and thoroughly evaluated its predictions using three complementary interpretability approaches: perturbation-based variable importance, value-dependent effect analysis, and leave-one-out (LOO) feature attribution. STraTS achieved the highest area under the precision--recall curve (AUPRC 0.171 {+/-} 0.022), compared with Random Forest (0.166 {+/-} 0.008), logistic regression (0.151 {+/-} 0.007), and XGBoost (0.128 {+/-} 0.010). Across interpretability methods, five predictors were consistently identified: birth weight, gender, Apgar score at 1 minute, umbilical serum thyroid stimulating hormone (uS-TSH), and treatment time in hospital. Indicators of early clinical severity, including chromosomal abnormalities and neonatal cerebral-status disturbances, showed the largest risk-increasing effects. Furthermore, the model's learned vector representations of subject-specific EHR sequences formed clinically coherent latent embeddings that reflect population heterogeneity along established perinatal risk dimensions. These findings demonstrate that combining multiple complementary interpretability methods yields stable, clinically plausible risk signals while revealing limitations that would remain undetected by any single approach, highlighting the importance of careful interpretability analysis of deep learning-based risk predictions.