Predicting Falls in the Nursing Homes via Recurrent Neural Network Models

Abstract

Background This study aimed to develop RNN algorithms to predict falls in nursing homes. Methods The MDS dataset and prescription drug exposure records were utilized to train RNN, LSTM, and GRU models for predicting falls within 90 days window. Results were compared to the previously evaluated CART-logit model. A ϕK correlation coefficient was used for feature analysis. Results RNNs performed similarly (AUROC ≈ 0.74). Feature analysis identified significant correlations for the delirium scale (ϕK = 0.63), use of antipsychotic medication (ϕK = 0.54), exposure to psychotropic medication (ϕK = 0.56), and cumulative number of days spent in the facility ( ϕK = 0.54). Conclusions All three models outperformed the CART-logit model, emphasizing significance of incorporating temporal aspects in fall prediction.