Joint prediction of chronic conditions onset : comparing multivariate probits with multiclass support vector machines

We consider the problem of building accurate models that can predict, in the short term (2–3 years), the onset of one or more chronic conditions at individual level. Five chronic conditions are considered: heart disease, stroke, diabetes, hypertension and cancer. Covariates for the models include standard demographic/socio-economic variables, risk factors and the presence of the chronic conditions at baseline. We compare two predictive models. The first model is the multivariate probit (MVP), chosen because it allows to model correlated outcome variables. The second model is the Multiclass Support Vector Machine (MSVM), a leading predictive method in machine learning. We use Australian data from the Social, Economic, and Environmental Factory (SEEF) study, a follow up to the 45 and Up Study survey, that contains two repeated observations of 60,000 individuals in NSW, over age 45. We find that MSVMs predictions have specificity rates similar to those of MVPs, but sensitivity rates that are on average 12% points larger than those of MVPs, translating in a large average improvement in sensitivity of 30%