The AI-Enabled Pharmacovigilance for Defence Health Surveillance: Automatic Detection of Adverse Drug Events from Patient Reviews Using BioClinical ModernBERT

Farmakovigilans Berbasis AI untuk Pengawasan Kesehatan Pertahanan: Deteksi Otomatis Kejadian Efek Samping Obat dari Ulasan Pasien Menggunakan BioClinical ModernBERT

Abstract

Pharmacovigilance is a critical component of post-marketing drug safety, yet conventional adverse drug event (ADE) reporting systems remain constrained by substantial underreporting. In defence health systems, the timely detection of medication-related safety signals is not only a clinical concern but also a matter of force health protection, medical readiness, and operational resilience. Patient-generated health narratives from online forums, drug review platforms, and social media provide a complementary source of pharmacovigilance intelligence, but their informal, unstructured, and context-dependent nature creates significant challenges for automated analysis. This study evaluates BioClinical ModernBERT, a biomedical–clinical long-context encoder based on the ModernBERT architecture, for automatic ADE detection from patient reviews. Its performance is compared with three representative BERT-based transformer baselines: BERT-base, BioBERT, and ClinicalBERT. Experiments were conducted using the CSIRO Adverse Drug Event Corpus (CADEC), a benchmark corpus of patient-reported medication experiences from online health forums. The corpus was pre-processed through sentence segmentation, text cleaning, medical entity normalization, and sentence-level label alignment for binary ADE/non-ADE classification. All models were fine-tuned using a 70:15:15 training, validation, and test split and evaluated using accuracy, precision, recall, and F1-score. The results show that BioClinical ModernBERT achieved the highest overall performance, with an F1-score of 0.891, outperforming ClinicalBERT (0.847), BioBERT (0.832), and BERT-base (0.798). Confusion matrix analysis further indicates that BioClinical ModernBERT reduced false negative errors, particularly in long, multi-clause, and clinically implicit patient narratives. These findings suggest that combining biomedical–clinical domain adaptation with long-context representation provides a meaningful advantage for detecting ADE signals in complex patient-generated text. From a defence health perspective, such capability may support the development of AI-enabled pharmacovigilance surveillance systems that enhance medication safety, health intelligence, and readiness-oriented risk monitoring across civilian–military health ecosystems.