Avaliação do desempenho de modelos de predição do risco de quedas relacionado aos medicamentos prescritos em adultos e idosos hospitalizados por meio de aprendizado de máquina

Abstract

INTRODUCTION: Falls may be one of the consequences of the use of risky drugs and/or drug interactions. Machine learning enables new approaches to risk stratisfication. Only one tool that assesses the risk of fall related to medication use was identified. Studies analyzing drugs as risk factors for falls through machine learning were not detected. OBJECTIVE: To evaluate the performance of models for predicting the risk of falls related to drugs prescribed in hospitalized adults and elderly people through machine learning. METHOD: Retrospective case-control study with adults and elderly people hospitalized at Hospital Nossa Senhora da Conceição in 2016. Age, prescribed drugs and drug classes were investigated. Data were exported to RStudio software for statistical analysis. The project was approved by the Scientific Committee of the School of Medicine of Pontifical Catholic University of Rio Grande do Sul and is linked to the umbrella project entitled “Detecção automática de eventos adversos utilizando processamento de linguagem natural nos prontuários eletrônicos de um hospital terciário”, approved by Research Ethics Committee. RESULTS: Prediction models developed through machine learning presented better performance when compared to an existing generalizable model. The models developed through gradient boosting algorithm, in general, presented better performance in relation to the others. The models that performed better in the population showed a decrease in performance when applied to the elderly subgroup. DISCUSSION: The present study proved that, in the study population, a model built from a dataset of a specific hospital presents better results in relation to a generalizable tool. Tools such as the Medication Fall Risk Score are restricted to a few variables, considering that health professionals must evaluate and calculate the score. Filling out these tools requires time and dedication from professionals, which could be applied in assistance. The models built through the gradient boosting algorithm stood out, with both drugs and drug classes variables. When applied to the elderly sample, the models built based on the population showed a decrease in performance. Therefore, it was decided to develop a specific model for this sample. CONCLUSION: Prediction models built through machine learning algorithms can help identify risk and improve patient care. The work of health professionals will not be replaced and the time spent on the application of scales can be directed to other important aspects of health care.