BioBERT-XGBoost for Adverse Drug Reaction Prediction: An Interpretable Hybrid Model for Risk-Aware Pharmacovigilance

Abstract

Adverse drug reactions (ADRs) are a critical challenge for patient safety, with over 21,000 alerts reported in Peru in 2024. Current artificial intelligence (AI) models in pharmacovigilance present limitations in external validation, clinical scalability, and algorithmic transparency. This work proposes BioBERT-XGBoost, an interpretable hybrid model that combines biomedical natural language processing with supervised machine learning to predict ADRs. The architecture integrates BioBERT for semantic extraction of pharmacological entities with XGBoost as a calibrated classifier, trained on public datasets (DrugBank, openFDA–FAERS) and anonymized clinical records. The pipeline includes standardized preprocessing through normalized vocabularies, feature engineering with semantic embeddings, class imbalance handling, and probability calibration. Evaluation uses discrimination metrics (AUROC, AUPRC), calibration (Brier score), and explainability (SHAP). The system is deployed on Microsoft Azure through a mobile application that generates risk-stratified clinical alerts, representing a step toward trustworthy clinical decision-support systems for proactive ADR detection.