Assessing the risk of falling in community-dwelling older adults through cognitive domains and machine learning techniques

Abstract

Background: Older people’s falls are a global public health problem, leading to injuries, disability, and fatalities. Using screening tools to measure predictive factors is essential for assessing the risk of falls among older adults. The literature highlights executive function tests as a way to assess this risk. They are also economical and reliable tools. Therefore, a Machine Learning (ML) technique based on variables obtained from cognitive domains could classify an older adult as at high or low risk of falling. Methodology: The study collected six variables from 50 community-dwelling older adults. The variables included age, educational level, and Trail Making Test (TMT) part B, Digital Span Backward, Stroop Color-Word Interference, and Mini Balance Evaluation Systems Test (Mini-BESTest) tests. These variables fed three ML models to predict if an older adult is at high or low risk of falling. Specifically, we considered Logistic Regression (LR), Decision Trees, and K-Nearest Neighbors. The proposed models were assessed using a bootstrapping sampling method and an aggregated confusion matrix, from which typical performance metrics were derived. The input variables in the best model were selected using a wrapper-based selection method. Results: Of the three models, the LR classifiers were top-ranked based on accuracy, with a maximum value of 71.4%. The best classifiers included the educational level or the TMT part B as input variables. Thus, these variables were strong predictors of fall events in the population study. We tested the input variables to ensure they were significant for the best LR classifiers and assessed model performance, generalization, and stability given the dataset sample size. Discussion: We weighed the performance metric results with a clinical perspective to select the best LR classifier. Thus, the more suitable model resulted in the classifier with TMT part B and educational level as input variables. Besides presenting competitive performance results, it enables us to consider a broader range of clinical information and draw more informed conclusions. Comparing our proposed model with four assessment tools, we observe it was second in Area Under the Receiver Operating Characteristic Curve (AUC) and third in accuracy. Conclusions: In this work, we developed an LR classifier to identify older adults with high or low risk of falling, using the TMT Part B test and the educational level as features. In addition, we provided cut-off values to assess the risk of falling using only the TMT part B test or the educational level. We found that, individually, 8 years or more of schooling or a result of the TMT part B lower than 212 s are associated, on average, with a low risk of falls. The Chilean health system can broadly implement the best classifier since the input variables are easy to collect, and the classification rule can be calculated using simple arithmetic operations.