(A) prediction model for prevention and management of metabolic syndrome based on machine learning

다국어 초록 (Multilingual Abstract)

Digital health-based lifestyle interventions (e.g., mobile applications, short message services, wearable devices, social media, and interactive websites) are widely used to manage metabolic syndrome (MetS). This study aimed to confirm the usefulness of digital health-based lifestyle interventions using healthcare devices and propose a novel prediction model of prevention and management for MetS. Participants with one or more MetS risk factors were recruited from December 2019 to September 2020, and finally, 106 participants were analyzed. Participants were provided with five healthcare devices and applications. Characteristics were compared at baseline and follow-up, and lifelog data that were collected during the clinical trial were analyzed. With these results, the frequency of use of healthcare devices for continuous self-care was quantified, and a novel prediction model for the prevention and management of MetS was developed. The model predicts persistence in continuous engagement as well as abbreviated risk factors for self-care effects. Representative machine-learning classifiers were used and compared. In both models, the random forest classifier showed the best performance, and feature selection was optimized through random forest-recursive feature elimination. As a result, the prediction model for persistence showed recall of 83.0%, precision of 92.4%, an F1-score of 0.874, a Matthews correlation coefficient (MCC) of 0.844, and accuracy of 94.9%. The prediction model for abbreviated risk factors showed a recall of 79.8%, a precision of 87.2%, an F1-score of 0.834, and an MCC of 0.797 for increased abbreviated risk factors, and a recall of 75.1%, a precision of 85.5%, an F1-score of 0.800, and an MCC of 0.747 for decreased abbreviated risk factors. The prediction model proposed showed high performance. Based on self-care with digital health-based lifestyle interventions, prediction models could be helpful for the prevention and management of MetS.