New Development of Statistical Machine Learning Methods with Applications to Large NHLBI Longitudinal Studies

• Activity Number: 249 - Making Impact for the Public Health: Innovative Methods and Challenges in Large NIH Studies
• Type: Invited
• Date/Time: Tuesday, August 4, 2020 : 1:00 PM to 2:50 PM
• Sponsor: ENAR
• Author(s): Colin O. Wu* and Xin Tian
• Companies: National Heart, Lung and Blood Institute, National Institutes of Health and National Heart, Lung and Blood Institute
• Keywords: Longitudinal analysis; dynamic disease prediction; Statistical machine learning; time-varying machine learning; functional predictor; landmark predictor

Abstract

Major epidemiological studies at the National Heart, Lung and Blood Institute (NHLBI) are almost exclusively longitudinal studies with long-term follow-up observations of risk factors, sub-clinical disease variables and disease events. Objectives of these studies include: (a) identifying important predictors for the prediction of disease events; (b) selecting useful “time-trends” and “landmarks” of influential variables to predict disease events; (c) developing statistical models for dynamic disease prediction and decision making. Traditional methods for longitudinal analysis fall short of (a) and (b). Existing machine learning methods fall short of (b) and (c). We present a series of statistical machine learning (ML) methods for correlated and dynamic feature selection and present their applications to some longitudinal studies at NHLBI. We show that these time-varying ML methods are capable of producing meaningful prediction models with “functional” and “landmark” longitudinal predictors. These NHLBI applications suggest that the effects of cardiovascular disease risks indeed change with time and should be treated as dynamic functions of appropriate time variables.