Deep Learning Models for Early Detection of Adverse Drug Reactions

Abstract

Adverse Drug Reactions (ADRs) remain one of the leading causes of preventable morbidity and hospitalization, highlighting the need for intelligent systems capable of identifying high-risk cases before clinical deterioration occurs. Existing pharmacovigilance approaches primarily depend on spontaneous reporting systems, which often detect ADRs only after significant patient harm has occurred. This study addresses the research gap by proposing a temporal deep learning framework that combines sequential medication histories, laboratory trends, and patient-specific clinical characteristics for proactive ADR prediction. Unlike conventional classification models that analyze isolated patient records, the proposed approach learns longitudinal clinical dependencies using a hybrid Bidirectional Long Short-Term Memory (BiLSTM) and attention mechanism. The framework is designed to improve prediction interpretability accuracy through while attention-based preserving feature importance. Experimental evaluation demonstrates consistent improvements over conventional machine learning baselines across multiple evaluation metrics. The findings indicate that integrating temporal clinical information substantially enhances early ADR detection and may support safer medication management in digital healthcare environments