Machine learning achieved 70.8% accuracy and 91.7% sensitivity in predicting physical activity decline in cardiac patients based on baseline digital health data.
Can machine learning applied to multimodal digital health data predict which patients will experience a decline in physical activity following a cardiac intervention?
Machine learning applied to baseline digital health and clinical data can accurately identify patients at high risk for physical activity decline after cardiac interventions, potentially enabling targeted early interventions.
Tasa de eventos absoluta: 0% vs 0%
Abstract Background Physical activity behaviour after cardiac interventions is a strong predictor of re-events and prognosis, yet identifying patients at risk for a decline of physical activity remains a challenge. Objective This study investigated whether machine learning applied to multimodal digital health data can predict which patients will reduce their physical activity following a cardiac intervention. Methods Eighty patients (mean age 73 (17) years, 89% male) were followed for 12 months after undergoing one of the following cardiac interventions: (a) coronary revascularisation (PCI or bypass surgery), (b) radiofrequency catheter ablation or electrophysiology study, or (c) transcatheter aortic valve implantation or valve surgery. The outcome of interest was future physical activity decline, defined as a sustained reduction of ≥1,000 steps/day from baseline during the follow-up period of one year. To predict this outcome, only data collected at baseline (immediately post-intervention) were used. Predictor variables included physical activity, sleep, heart rate, circadian rhythm metrics, lifestyle behaviour, clinical, psychosocial, and demographical patient characteristics. A tree-bagging ensemble classifier was trained using Bayesian optimization with five-fold cross-validation. Model performance was evaluated on a held-out 30% test set using step-count–based outcome labels. Results The final model achieved 70.8% accuracy, with high sensitivity (91.7%) for predicting physical activity decline and an AUC of 0.833. Feature importance analysis revealed that age, circadian rhythm amplitude, exertion scores, activity counts, and cardiorespiratory fitness were the strongest predictors of decline. Patients with low baseline activity and disrupted rhythms were at highest risk. Conclusion Machine learning can identify high-risk profiles for physical activity decline using data collected shortly after cardiac intervention. Integrating such models into digital health platforms could enable early, targeted interventions and support better informed and more personalized cardiac care.
Blerck et al. (Thu,) reported a other. Machine learning achieved 70.8% accuracy and 91.7% sensitivity in predicting physical activity decline in cardiac patients based on baseline digital health data.