Patient-Specific Modeling with Personalized Decision Paths

Abstract

Predictive models can be useful in predicting patient outcomes under uncertainty. Many algorithms employ "population" methods, which optimize a single model to perform well on average over an entire population, but the model may perform poorly on some patients. Personalized methods optimize predictive performance for each patient by tailoring the model to the individual. We present a new personalized method based on decision trees: the Personalized Decision Path using a Bayesian score (PDP-Bay). Performance on eight synthetic, genomic, and clinical datasets was compared to that of decision trees and a previously described personalized decision path method in terms of area under the ROC curve (AUC) and expected calibration error (ECE). Model complexity was measured by average path length. The PDP-Bay model outperformed the decision tree in terms of both AUC and ECE. The results support the conclusion that personalization may achieve better predictive performance and produce simpler models than population approaches.

Figure 1:. An example decision tree and a personalized decision path for predicting in-hospital mortality for a patient admitted with heart failure. Panel (a) lists the variables and corresponding values for a patient (test case) whose outcome we want to predict. Panel (b) shows a decision tree and the path (arrows in bold) used for inference for the patient, and panel (c) shows a personalized decision path (derived by the PDP-Bay method that is described in the section on Algorithmic Methods) for the patient. Terminal leaf nodes contain counts of corresponding training samples who survived or died during hospitalization. The patient in this case survived.

Figure 2:. Pseudocode for the PDP-Bay method.

Figure 3:. ROC plots for DT, PDP-Ent, and PDP-Bay methods applied to the chronic pancreatitis dataset. The respective AUCs and estimated 95% confidence intervals are also provided. Note that the confidence intervals of the PDP-Bay and DT do not overlap, indicating statistically significantly better performance of the personalized method.