Mixed-Effect Modeling for Longitudinal Prediction of Cancer Tumor

Abstract

In this paper, a mixed-effect modeling scheme is proposed to construct a predictor for different features of cancer tumor. For this purpose, a set of features is extracted from two groups of patients with the same type of cancer but with two medical outcome: 1) survived and 2) passed away. The goal is to build different models for the two groups, where in each group, patient-specified behavior of individuals can be characterized. These models are then used as predictors to forecast future state of patients with a given history or initial state. To this end, a leave-on-out cross validation method is used to measure the prediction accuracy of each patient-specified model. Experiments show that compared to fixed-effect modeling (regression), mixed-effect modeling has a superior performance on some of the extracted features and similar or worse performance on the others.