Predicting recreational therapy engagement in veterans’ long-term care: a machine learning approach

Background There is limited understanding regarding the factors that predict recreational therapy engagement among veterans in long-term care facilities. We aimed to develop and validate machine learning models to predict recreational therapy participation and identify key factors influencing engagement patterns among veterans in long-term care facilities. Methods In this cross-sectional observational study, we used data from 57 veterans aged 18 years and above at the New York State Veterans Home at Oxford. Data were collected through a comprehensive self-administered survey capturing demographic characteristics, participation patterns, and activity preferences. Two binary outcome variables were constructed: high participation and any participation. Five machine learning algorithms (Random Forest, Decision Tree, Gradient Boosting, and Logistic Regression with L1 and L2 regularization) were systematically evaluated using Leave-One-Out Cross-Validation for high participation and 5-fold Stratified Cross-Validation for any participation. Feature selection was implemented using SelectKBest with f_classif scoring, and class imbalance was addressed through balanced weighting techniques. Results Random Forest emerged as the optimal algorithm for both prediction tasks, achieving F1-scores of 0.860 ± 0.347 for high participation prediction and balanced accuracy of 0.619 ± 0.081 for any participation prediction. Feature importance analysis revealed activity preference diversity (Gini importance: 0.293) and total preferences (0.254) as the primary predictors of high participation, while facility tenure (0.268) was the strongest predictor of any participation. Veterans with preference diversity >4.5 activities combined with satisfaction scores >3.84 achieved 100% observed probability of sustained high participation [ n = 5; 95% exact binomial CI: (47.8%, 100%)], though this estimate should be interpreted cautiously given the small subgroup size. New residents (≤1.5 years) with limited preferences demonstrated the highest risk for non-participation. Group activities (Gini importance: 0.143) and spiritual activities (Gini importance: 0.100–0.101) emerged as significant predictors across both models. Conclusions This research provides the first proof-of-concept demonstration of a machine learning approach for predicting recreational therapy engagement among veterans in long-term care facilities, establishing methodological feasibility and generating testable hypotheses for prospective multi-site validation. Activity preference diversity and facility tenure serve as primary determinants of participation, with a critical 1.5-year adaptation period identified for intervention targeting. These predictive models can be applied during admission or early in residency to identify veterans at risk of low participation, enabling recreational therapy staff to implement tailored, proactive engagement strategies before disengagement occurs.