Abstract P274: Beyond MVPA: Machine Learning Insights Into Activity Patterns and Cardiometabolic Health in the National Health and Nutrition Examination Survey (NHANES) 2003-2004

Abstract

Physical activity (PA) and sedentary behavior (SB) volumes and patterns may be useful in identifying individuals who are at elevated cardiometabolic risk. Identifying individuals at risk will facilitate early interventions for prevention of cardiovascular diseases. The objective of this project is to identify important free-living physical activity patterns to classify cardiometabolic markers including insulin, plasma glucose, and triglyceride. Methods: We used cross-sectional NHANES 2003-2004 accelerometer, blood exam and demographic data. We included adults age ≥18 years who wore a hip-worn ActiGraph 7164 during waking hours for at least 4 days of 10 hours/day of wear time. We included PA and SB predictors including minutes per day (min/d) in sedentary time (0-99 counts per minute, cpm), light intensity PA (100-759 cpm), Lifestyle PA (760-2019 cpm), moderate to vigorous PA (MVPA, ≥2019 cpm), number of breaks in SB, and number of 30-min and 60-min SB, 10-min MVPA bouts etc. The cut points for cardiovascular disease risk were defined based on literature as: fasting serum insulin ≤ 5 μU/mL, plasma glucose ≤ 100 mg/dL and triglyceride < 150 mg/dL. Three random forest (RF) classification models were developed for each cardiometabolic marker using age, body-mass-index, general health, and PA variables as predictors. Models were trained and tested using a 60% training - 40% testing split. We tuned the hyperparameters (final selection: number of trees = 500 and variables = 3) to find the model with the highest testing accuracy based on the out-of-bag error. The importance of each predictor was measured and ranked by the mean decrease in GINI index. Results: The three datasets include insulin (n = 2810, age=34 ±23, 50% females), glucose (n = 1852, age = 49±20, 51% females) and triglyceride (n = 3180, age=36 ±24, 50% females). For insulin model, the PA variables with largest mean decrease in GINI index were SB time, light intensity PA, MVPA, and SB bouts of 30-min. Similarly, time spent in Lifestyle PA, light intensity PA, number of SB breaks and SB bouts of 30-min were the most important variables for abnormal fasting glucose prediction. The number of sedentary breaks, light intensity PA, SB time and SB bouts of 30-min were the most important PA predictors for risk of high triglyceride. Both BMI and age were important demographic variables for prediction of increased risk in each model. The prediction accuracy and precision are as following: Insulin(73%; 76%), Plasma Glucose(68%; 58%), and Triglyceride(64%; 46%). Conclusion: Cardiometabolic markers using free-living PA has the potential to provide insights for classifying the risk for cardiovascular disease based on given physiological marker cut-off points. The time spent lower intensity PA patterns and the number of sedentary breaks, and 30-minutes bouts are important predictors for identifying individuals who may have increased cardiometabolic risk.