Developing a deep learning natural language processing algorithm for automated reporting of adverse drug reactions

The detection of adverse drug reactions (ADRs) is critical to our understanding of the safety and risk-benefit profile of medications. With an incidence that has not changed over the last 30 years, ADRs are a significant source of patient morbidity, responsible for 5%-10% of acute care hospital admissions worldwide. Spontaneous reporting of ADRs has long been the standard method of reporting, however this approach is known to have high rates of under-reporting, a problem that limits pharmacovigilance efforts. Automated ADR reporting presents an alternative pathway to increase reporting rates, although this may be limited by over-reporting of other drug-related adverse events. We developed a deep learning natural language processing algorithm to identify ADRs in discharge summaries at a single academic hospital centre. Our model was developed in two stages: first, a pre-trained model (DeBERTa) was further pre-trained on 1.1 million unlabelled clinical documents; secondly, this model was fine-tuned to detect ADR mentions in a corpus of 861 annotated discharge summaries. This model was compared to a version without the pre-training step, and a previously published RoBERTa model pretrained on MIMIC III, which has demonstrated strong performance on other pharmacovigilance tasks. To ensure that our algorithm could differentiate ADRs from other drug-related adverse events, the annotated corpus was enriched for both validated ADR reports and confounding drug-related adverse events using. The final model demonstrated good performance with a ROC-AUC of 0.955 (95% CI 0.933 - 0.978) for the task of identifying discharge summaries containing ADR mentions, significantly outperforming the two comparator models.