Detecting Chronic Diseases from Sleep-Wake Behaviour and Clinical Features

Many chronic diseases show evidence of correlations with sleep-wake behaviour, and there is an increasing interest in making use of such correlations for early warning systems. This research presents an approach towards early chronic disease detection by mining sleep-wake measurements using deep learning. Specifically, a Long-Short-Term-Memory network is applied on actigraph data enriched with clinical history of patients. Experiments and analysis are performed targeting detection at an early and advanced disease stage based on different clinical data features. The results show for disease detection an averaged accuracy of 0.62, 0.73, 0.81, 0.77 for hypertension, diabetes, sleep apnea and chronic kidney disease, respectively. Early detection performs with an averaged accuracy of 0.49 for sleep apnea and 0.56 for diabetes. Nevertheless, compared to existing work, our approach shows an improvement in performance and demonstrates that predicting chronic diseases from sleep-wake behavior is feasible, though further investigation will be needed for early prediction.