Predicting sleep disturbance and healthcare utilisation using survey data

Abstract

Objective: To evaluate the ability of machine learning classification models to predict sleep disturbance and high healthcare utilisation among older adults using categorical national health survey data. Method and Material: A cross-sectional observational study was conducted using de-identified survey data from 714 adults aged 65–80 years. Multiple machine learning algorithms were implemented in the Waikato Environment for Knowledge Analysis (WEKA) and assessed using stratified ten-fold cross-validation. Performance was evaluated using accuracy, sensitivity, specificity, precision, F1-score, and area under the receiver operating characteristic (ROC) curve (AUC). Results: All evaluated models achieved predictive accuracy above 70% for both outcomes, with generally strong sensitivity, F1-score, and predictive performance. Ensemble-based approaches, particularly gradient boosting and random forest, demonstrated the strongest overall performance across accuracy, AUC, and F1-score metrics. Logistic regression and support vector machines also showed stable and competitive results across outcomes, balancing prediction accuracy with relatively consistent sensitivity and specificity. Conclusions: Routinely collected survey data can be effectively used to identify older adults at increased risk of sleep problems and high healthcare utilisation. Machine learning models may support preventive interventions and more efficient healthcare planning.