Speaker: Jeong Hoon Jang, Yonsei University "A functional quantile regression model for predicting hypoglycemic events using continuous glucose monitoring data"

Location: Zoom

Abstract: Continuous glucose monitoring (CGM) with wearable devices is becoming the standard of diabetes care. Models have been developed to depict the temporal patterns of glucose variation, but few models have been developed to predict hypoglycemic events---sudden and dangerous drops of glucose in the blood. A methodological challenge in predicting hypoglycemic events is that the common statistical models for mean outcomes do not capture the glucose variation in the lower quantiles. In this research, we describe a function-on-function quantile regression model for assessing the risk of hypoglycemia. By modeling the lower quantiles, we provide short-term predictions for glucose variation in the lower quantiles and identify patient factors that are associated with the risk of such events. Furthermore, to take into account various distinct patterns of glucose curves which may not be well predicted by a single model, we combine the proposed functional quantile regression approach with probabilistic functional classification. Specifically, we allow regression parameters to depend on unobserved class memberships of predictor curves probabilistically determined according to their distinct patterns, thereby flexibly incorporating heterogeneous predictor-response relationships and enhancing predictive accuracy. We developed a Bayesian computational approach for parameter estimation. Extensive simulation studies showed that the model has excellent short-term predictive performance. In putting forward this functional regression model, we hope to provide an example on predicting clinically important events by using wearable device data.