Abstract Body (Do not enter title and authors here): Background: Behavioral factors such as physical activity, sleep, and heart rate are associated with cardiovascular (CV) health but are not used in current CV risk prediction tools. Consumer wearable devices offer continuous behavioral data that may enhance atherosclerotic cardiovascular disease (ASCVD) risk prediction.
Hypothesis: Incorporation of wearable-derived metrics into CV risk prediction algorithms may improve accuracy. Development of demographic and wearable-only CV risk prediction algorithms may enable continuous and more scalable monitoring compared to existing models that require in-person assessments.
Methods: We conducted a retrospective cohort study using electronic health record (EHR) data from the All of Us Research Program linked to participants’ Fitbit data. Adults (≥18 years) with EHR and Fitbit data (up to 180 days prior to ASCVD event) were included; those with pre-existing CV disease were excluded. Incident ASCVD was defined using ICD-10 codes (myocardial infarction, ischemic stroke) and CPT codes (percutaneous coronary intervention and coronary artery bypass grafting). ASCVD risk was assessed using PREVENT equations. Multiple imputation was used for missing data (up to 2 missing variables per participant). Logistic regression was used to estimate ASCVD risk. Model performance was assessed using the area under the receiver operating characteristic curve (AUC-ROC) and Youden's J statistic.
Results: Among 4,193 participants (median age 52, 74% female) with a calculable PREVENT score, 162 had incident ASCVD during a median follow-up of 4.1 years. Those with incident ASCVD were older (59.4 vs. 51.2 years, P<0.001), more often male (47% vs. 25%), and had higher rates of hypertension (47% vs. 37%, P=0.01) and smoking (48% vs. 33%, P<0.001). A wearable-based model supplemented with basic, patient-known demographic information achieved an AUC of 0.741 (Table 1) vs. AUC of 0.728 for PREVENT alone. Adding wearable data to PREVENT yielded modest improvement (AUC 0.738 vs 0.728), but with low specificity. A wearable-based model with demographics predicted the ASCVD risk category (≥7.5% vs. <7.5% risk per PREVENT) with an AUC 0.95, 77% sensitivity, and 95% specificity (Table 2).
Conclusions: Wearable-based models using behavioral metrics and basic demographics perform similarly to PREVENT, and offer remote, personalized, dynamic ASCVD risk assessment. Model adjustments to improve specificity and validation in other cohorts are warranted.