Toward personalized exergaming: a machine learning-based decision-support approach to cognitive-motor training for long-term care residents with cognitive impairments

Abstract Background Major Neurocognitive Disorder (MND) is characterized by progressive impairments across multiple cognitive domains, leading to reduced autonomy, diminished quality of life, and frequent transition to long-term care facilities. Cognitive-motor exergames, which combine full-body movements with embedded cognitive tasks, represent a promising intervention to slow down the declines in physical and cognitive functioning and maintain engagement. However, personalized implementation strategies remain largely unexplored. This study aimed to develop a machine-learning–based approach to generate personalized initial game recommendations and tailored assistance needs for residents of long-term care facilities with cognitive impairments. Methods A cross-sectional study was conducted with 100 residents. Demographic characteristics, cognitive function assessed with the Quick Mild Cognitive Impairment (QMCI) screen, and motor function assessed with the Short Physical Performance Battery (SPPB) were recorded. Participants completed 20 exergames on the Dividat Senso system, and gameplay ability with required verbal and physical assistance was systematically documented. A LightGBM model was trained to predict gameplay ability and assistance needs based on individual characteristics and individual game identity. Results Participants were on average 86.6 years old and predominantly female (71%). As expected, they exhibited substantial cognitive and motor impairments. The machine-learning model achieved a macro-averaged area under the receiver operating characteristic curve (ROC-AUC) of 0.79 in cross-validation and 0.77 in the held-out test set. Explainability analyses highlighted that individual game and QMCI scores were the strongest contributors to model predictions. While the model demonstrated robust discrimination of gameplay ability, predicting assistance needs showed lower classification performance (AUC = 0.67–0.71). Conclusions These findings demonstrate the potential of clinically accessible data to support personalized, data-driven cognitive-motor training recommendations for long-term care residents with cognitive impairment. Individual game and cognitive status emerged as stronger determinants of gameplay ability than physical or demographic factors, underscoring the importance of cognitively informed game selection. Our decision-support model showed promising internally validated discrimination and strong potential for personalized game selection, whereas larger and more diverse datasets are needed to optimize individualized support recommendations.