Local multivariate modeling for predicting body composition indifferent segmental compartments

abstracts dans le "Final Programme" onlineInternational audienceBody composition is an important indicator for public health issues.The two main components, body fat and body lean masses, enable to assess nutritional and physio-pathological status of individuals. Body fat, particularly abdominal fat mass, is associated with anincrease in cardio-metabolic morbidity and mortality. The decrease in body lean masses during aging is linked with loss of autonomy. The ability to predict body composition from simple anthropometriccovariates such as gender, age, height, weight and waist circumference, is an important issue. The aim is not to replace measurement methods (DXA), but to propose a pre-diagnostic tool todetermine whether a DXA measurement is required or not. A locally weighted multivariate regression was proposed to predict simultaneously several compartments. The aim of the approach isto build a good prediction from a reference dataset (NHANES).Among all tried methods, after applying the models on differents datasets, it turns out that the locally weighted approach gives the more reliable prediction, also comparing to published methods