Identification of adverse drug reactions in Dutch electronic health records

Abstract

Older people with polypharmacy and multimorbidity are at high risk for adverse drug reactions (ADRs) due to drug-drug interactions and age-related changes in pharmacodynamics/kinetics. A significant number of ADRs are due to represcription of withdrawn medication: recurrent ADRs. ADRs are poorly documented and are often lost in electronic health records (EHRs) in the form of unstructured texts.

Therefore we developed a method to detect potential ADRs in unstructured text of EHRs. For this we applied MedCAT, a concept extraction and linking tool, to the admission and discharge letters of 93 acutely admitted geriatric patients with polypharmacy. These letters were previously screened by clinicians for the presence of triggers and drugs frequently causing ADRs. We used MedCAT to extract text fragments containing adverse drug events (ADEs) frequently associated with ADRs and labelled these for whether the clinician denotes an ADR. Then, we used these texts to train two BERT-based models to recognize ADRs mentioned by clinicians in texts. These models were evaluated in terms of precision, recall and f1 score.

Our strategy to recognize ADEs frequently associated with ADRs achieves an f1 score of 71.4%, also detecting new ADEs that were missed during screening in previous research. The best model for recognizing ADRs mentioned by clinicians yields an f1 score of 76.9% and manages to outperform the baseline by 18.0%.

We argue that our method performs well at identifying ADEs frequently associated with ADRs. Our strategy for identification of potential ADRs should however be further trained (on a larger dataset) and optimized before it can play a role in preventing recurrent ADRs in clinical practice.