UC Berkeley

We sought to develop an improved decision algorithm for applications in precision medicine. This algorithm should perform as well or better than existing prognostic scores, be interpretable and parsimonious in the predictors it uses for an individual patient, and be able to perform competitively against best-in-class existing ensemble learning algorithms. The algorithm must be flexible enough to predict well in heterogeneous patient populations, where we might expect the covariates that are related to the outcome of interest to be different for different types of patients. To this end, a new supervised classification method is proposed which performs both dimension and instance reduction data-adaptively to hone in on only the most relevant information for a given patient.