Machine learning applied to wearable fitness tracker data and the risk of hospitalizations and cardiovascular events

Abstract

Background: Wearable fitness trackers generate extensive physiological and activity data, offering potential to monitor health and predict outcomes. Machine learning (ML) techniques applied to these data may enable early identification of adverse health conditions, such as hospitalizations and development of cardiovascular diseases (CVD). This study aimed to evaluate ML models' ability to forecast the incidence of (1) hospitalizations from any cause and (2) of new diagnosis of CVD, including a composite of heart failure (HF), coronary artery disease or myocardial infarction (CAD-MI), cardiomyopathy (CMP), and atrial fibrillation (AF). Method and Results: Data from 14,157 participants in the All of Us study that included both Fitbit and electronic health record (EHR) information were censored on the date preceding events and analyzed using various ML classifiers for extracted feature data. Performance metrics included accuracy, area under the receiver operating characteristic (AUROC) curve, and F1 scores. Our overall study population was young (median age 54 years), with good representation of women (67%). For hospitalizations, a Random Forest classifier achieved the best performance (AUROC=0.95, accuracy=0.99, F1 score=0.92). For the CVD events, the best prediction model was gradient boosting (AUROC=0.80, accuracy=0.71, F1 score=0.15).

Conclusion: ML models applied to Fitbit data demonstrate promise in predicting clinical outcomes with strong performance for predicting all-cause hospitalizations and modest performance for predicting incident CVD. Wearable technology could play a role in risk assessment and patient management.