Multi-method statistical signal aggregation with machine learning for severity classification of neonatal adverse drug reactions.

Neonates represent one of the most pharmacologically vulnerable patient populations, yet they remain systematically underrepresented in clinical drug safety trials. Spontaneous adverse drug reaction (ADR) reports submitted to the FDA Adverse Event Reporting System (FAERS) constitute the primary real-world evidence base for neonatal pharmacovigilance. The multi-drug and multi-reaction structure of raw FAERS reports, combined with the absence of standardised severity labels, has precluded the development of data-driven severity prediction for this population. This work presents an end-to-end computational pipeline that integrates four-method pharmacovigilance signal detection with supervised ADR severity classification for neonates and empirically investigates whether disproportionality signal scores can augment machine learning classifiers beyond standard demographic and co-occurrence features. A neonatal-specific dataset of 826,516 individual drug-ADR association records was constructed from 32,441 raw FAERS reports through systematic drug ingredient splitting and reaction disaggregation, where each report contributes one record per unique drug-ADR combination, with a five-tier severity hierarchy (fatal, critical, serious, moderate, non-serious) derived from FAERS outcome flags. Pharmacovigilance signal detection was conducted using four statistical methods and a comprehensive feature ablation analysis evaluated three feature sets (FS) base demographic (FS1), signal-only (FS2), and combined (FS3) across six classifiers, namely logistic regression, random forest, extra trees, histogram gradient boosting, AdaBoost, and multi-layer perceptron, on both binary and multiclass severity tasks. Of 124,841 unique drug-ADR pairs in the expanded dataset, 34,698 (27.8%) had sufficient co-occurrence reports (n ≥ 3) to be evaluated for pharmacovigilance signals. A total of 13,506 pairs achieved all-four-method consensus, representing 38.9% of evaluated pairs and 10.8% of all unique pairs, spanning 792 drugs and 1778 ADR terms. Histogram gradient boosting achieved the strongest performance on the binary task (AUROC = 0.906, AUPRC = 0.954) and the multiclass task (AUROC = 0.874, AUPRC = 0.579) for severity classification in order to establish a robust benchmark for data-driven neonatal ADR severity prediction.