Predicting fall risks with machine learning and medication data

Patient falls present a costly issue burdening nursing staff and patient safety alike. In response, the accurate prediction of fall risks has been previously established as a suitable application field for machine learning. Previous studies have shown promise, but were oftentimes limited in scope, cohort sizes and feature variety. Especially complex medication patterns have been insufficiently investigated for this purpose. This work developed machine learning (ML) models with a focus on evaluating the predictive potential of medication data. Performance was assessed and compared against a manual fall risk assessment (FRA) currently employed at Charité. The ML models outperformed the manual FRA on precision but were slightly outperformed on recall. In parallel, using exploratory statistics, individual medications with a strong association to fall risks were identified and subsequently presented to clinical experts for evaluation. This work could inform decision makers about the advantages and disadvantages of substituting manual FRA with ML models and contribute to clinical guidelines on medication administration. Ultimately, if pursued adequately, the findings have the potential to alleviate nursing staff workload and improve patient outcomes.