Semi-supervised neighborhoods and localized patient outcome prediction

Robust statistical methods that can provide patients and their healthcare providers with individual predictions are needed to help guide informed medical decisions. Ideally an individual prediction would display the full range of possible outcomes (full predictive distribution), would be obtained with a user-specified level of precision, and would be minimally reliant on statistical model assumptions. We propose a novel method that satisfies each of these criteria via the semi-supervised creation of an axis-parallel covariate neighborhood constructed around a given point defining the patient of interest. We then provide non-parametric estimates of the outcome distribution for the subset of subjects in this neighborhood, which we refer to as a localized prediction. We implement local prediction methods using dynamic graphical methods which allow the user to vary key options such as the choice of the variables defining the neighborhood, and the size of the neighborhood.