Integrating Clinical Modeling and Machine Learning for Risk Assessment of Paracetamol and Other Nonsteroidal Anti-Inflammatory Drug Hypersensitivity in Children.

Abstract

Nonsteroidal anti-inflammatory drug (NSAID) hypersensitivity is a common cause of drug-related reactions in children. Pretest risk stratification may improve the safety and efficiency of drug provocation testing. To develop a clinically interpretable risk stratification tool (nomogram + simplified score) for pediatric paracetamol and/or other NSAID hypersensitivity and to validate its performance against machine learning (ML) models. We conducted a retrospective cohort study (2014-2025) of children evaluated for suspected paracetamol and/or other NSAID hypersensitivity. Analyses used the index reaction as the unit, classifying definitive outcomes as NSAID-hypersensitive or NSAID-tolerant. Independent predictors from multivariable logistic regression were used to develop a clinically interpretable risk stratification tool, implemented as a nomogram and a simplified point-based score. were trained. Eight ML models were trained using fivefold cross-validation under three data scenarios (original, matched, and Synthetic Minority Oversampling Technique for Nominal and Continuous Variables). Among 507 index reactions (from 487 children) evaluated for suspected paracetamol and/or other NSAID hypersensitivity, 90 of 507 (17.7%) had confirmed hypersensitivity. Independent predictors were age 82.5 months or older at the time of reaction, coexisting asthma and/or allergic rhinitis, latency between exposure and symptom onset of 60 minutes or less, having angioedema, respiratory symptoms, and hypotension or syncope during the index reaction. The nomogram and simplified point-based score showed strong discrimination (receiver operating characteristic [ROC] area under the curve [AUC] = 0.877) and bedside applicability. After class balancing (Synthetic Minority Oversampling Technique for Nominal and Continuous Variables), ensemble ML achieved top performance: gradient boosting ROC AUC = 0.955, recall = 0.895, and F1 = 0.896; random forest ROC AUC = 0.953, recall = 0.890, and F1 = 0.883; and AdaBoost ROC AUC = 0.940, recall = 0.873, and F1 = 0.874. The nomogram and simplified point-based score provide practical pre-drug provocation testing risk stratification for children evaluated for suspected paracetamol and/or other NSAID hypersensitivity. Ensemble ML can complement the tool by improving sensitivity to minimize false negatives. Multicenter external validation and prospective impact studies are warranted before clinical implementation.