ENHANCED ADVERSE-EVENT DETECTION AND DRUG-EVENT RELATION EXTRACTION FROM CLINICAL NOTES

Abstract

Adverse drug events (ADEs) are a significant source of preventable patient harm, yet many ADE signals remain buried in free-text clinical notes. Clinical notes often describe adverse events (AEs) in two ways: a drug may cause the event, in which case the event represents an ADE, or a drug may be given to treat the event, in which case the event represents the Reason for treatment. In the 2018 N2C2 Track 2 benchmark, ADEs and Reasons are annotated as separate entity types, despite often being similar in both wording and clinical meaning. This similarity makes them difficult to distinguish during named entity recognition (NER), leading to missed event spans and downstream errors in relation classification. This dissertation investigates whether ADE and Reason mentions are better handled as a unified adverse event category during concept extraction, with the final ADE-versus-Reason distinction resolved at the relation classification stage. First, transformer-based BiLSTM-CRF models were evaluated using the original N2C2 annotation scheme. The results showed strong performance for structured medication-related entities, such as Drug, Strength, Route, Frequency, Form, Dosage, and Duration. However, ADE and Reason remained the most difficult entity types. The best model, GatorTron-BiLSTM-CRF, achieved a lenient micro-F1 of 0.9453, but ADE and Reason F1 scores were lower, at 0.6769 and 0.7538, respectively. These findings support the hypothesis that ADE and Reason are not ideal as separate entity labels during NER because their interpretation depends heavily on drug-event context. To address this limitation, this dissertation proposes an AE-first two-stage framework. The first stage detects AEs as a unified event category, including symptoms, diseases, abnormal findings, infections, abnormal laboratory values, and abnormal vital signs. The second stage generates Drug-AE candidate pairs and classifies each pair as Drug-ADE, Drug-Reason, or No-Relation. In end-to-end evaluation on the N2C2 2018 benchmark, the proposed framework achieved F1 scores of 0.93 for Drug-ADE and 0.94 for Drug-Reason using LLM-based AE detection and a T5-EHR relation classifier. These results improved over previously reported end-to-end benchmarks of 0.75 for Drug-ADE and 0.84 for Drug-Reason, supporting a more precise task formulation in which AEs are detected broadly first and the ADE-versus-Reason distinction is resolved at the relation layer. Furthermore, this dissertation contributes a manually annotated AE dataset developed from the N2C2 2018 corpus. The dataset maps original ADE and Reason mentions into a unified AE category and adds additional AE mentions that were not annotated in the original relation-centered scheme. These added mentions include diseases, symptoms, abnormal findings, infections, abnormal laboratory values, and abnormal vital signs, regardless of whether they are linked to a drug. A QuickUMLS analysis showed substantial semantic overlap among the original ADE, the original Reason, and manually added AE mentions, supporting the use of a unified AE category. A GatorTron-BiLSTM-CRF model trained on this manually annotated dataset achieved an AE lenient F1 of 0.88 and detected 97.0% of original ADE spans and 95.0% of original Reason spans as AEs. Finally, the manually trained AE detector was integrated into the end-to-end pipeline. This supervised AE-first pipeline achieved F1 scores of 0.93 for Drug-ADE and 0.96 for Drug-Reason, showing that the AE-first formulation can be implemented without relying on LLM-generated spans. All in all, this dissertation demonstrates that ADE and Reason extraction should not be treated only as a standard entity classification problem. Detecting adverse events broadly before relation classification reduces upstream error propagation, improves end-to-end drug-event extraction, and provides a foundation for more reliable pharmacovigilance-oriented clinical NLP systems.