A Bidirectional Encoder Representations from Transformers-Based Framework for Automated Adverse Drug Reaction Detection from Patient-Generated Psychiatric Treatment Narratives

Adverse drug reactions (ADRs) associated with psychiatric medications are difficult to identify because patient-reported experiences are often informal, diverse, and embedded in unstructured textual narratives. Conventional pharmacovigilance systems mainly depend on structured reports, which may overlook valuable safety information contained in real-world patient communications. This study develops a transformer-based natural language processing (NLP) framework using a fine-tuned Bidirectional Encoder Representations from Transformers (BERT) model for automated ADR detection in psychiatric treatment narratives. Two publicly available datasets, Psychiatric Treatment Adverse Reactions (PsyTAR) and Consumer Adverse Drug Event Corpus (CADEC), were integrated to construct a domain-specific corpus for binary classification of ADR and non-ADR texts. The preprocessing pipeline included text normalization, BERT-based tokenization, class balancing, and supervised fine-tuning using the HuggingFace Transformers framework. The proposed model was trained with the BERT-base-uncased architecture and evaluated using standard classification metrics on the validation dataset. Experimental results demonstrate that the model achieved an F1-score of 0.89, accuracy of 0.89, precision of 0.88, and recall of 0.91, indicating its effectiveness in identifying ADR-related expressions from patient-generated narratives. The findings suggest that transformer-based models can support automated pharmacovigilance workflows by extracting safety signals from unstructured psychiatric healthcare texts. However, further investigation using larger clinical datasets, domain-specific pretrained models, and explainable artificial intelligence methods is required before practical clinical deployment.