Predicting risk of falling in older adults using supervised machine learning: a comparative analysis of model performance.

Abstract

Background: Falls in older adults pose a significant health risk and reliable predictive models for assessing the risk of falling would be highly beneficial and clinically relevant. This study evaluates the performance of various supervised machine learning (ML) algorithms in predicting the risk of falling (≥ 1 self-reported fall in the past year) using balance and functional ability measures.

Methods: Data from 94 older adults were analyzed incorporating comprehensive assessments of physical function and balance, including the five-repetition sit-to-stand test (5XSTS), 30-second chair stand test (30CST), Berg balance scale (BBS), hip abduction strength, 6‑minute walk test (6 MWT) and 10-meter walk test (10 MWT). We implemented and compared four ML models: partial least squares discriminant analysis (PLS-DA), linear discriminant analysis (LDA), support vector machines (SVM) and k‑nearest neighbors (k-NN). Model performance was evaluated using cross-validation, with sensitivity, specificity, precision and accuracy. To provide a clinically interpretable benchmark, stepwise logistic regression with cross-validation was also applied.

Results: All ML models demonstrated strong discriminatory power. The PLS-DA achieved the highest sensitivity (0.96), specificity (0.96), precision (0.96), accuracy (0.96) and area under the receiver operating characteristic curve (AUC, 0.97). The LDA and k‑NN exhibited balanced overall AUCs (0.94 and 0.96, respectively). The regression benchmark consistently retained a small subset of predictors, most often the 6 MWT, 30CST and BBS. These models achieved mean accuracy of 0.84, sensitivity of 0.82, specificity of 0.85 of and AUC of 0.94.

Conclusion: Supervised ML models effectively predict the risk of falling in older adults, with PLS-DA emerging as the most robust classifier. While SVM showed strong predictive power, other models provided better clinical interpretability. Regression benchmarks highlight that a few functional measures already perform strongly but ML further improves classification by integrating multidimensional patterns. Importantly, as our design is retrospective, these findings represent a classification of the fall history rather than definitive prediction of future falls.