Development of Prediction Models using Clustering

Abstract

Repeated measures data have been commonly generated in many clinical studies. One of main objectives of the repeated measures data analysis is to predict a future outcome from the previously observed values. Repeated measures data often show certain patterns over time which can be easily checked by simple scatter plots. In this paper, we demonstrate that the use pattern information increases the accuracy of prediction in repeated measures data analysis. We propose to make a prediction model first by clustering data patterns via clustering methods and later by adding this clustering information into our model as a variable. We illustrate our approach using a real clinical data for bipolar patients. One of the clinical outcomes is Clinical Global Impression (CGI) value to predict patients the extent of depression. We chose the better measure from calculated distances, Euclidean and 1-corr, between individual CGI values and clustered them by hierarchical clustering methods. Then we developed best prediction model of the extent of depression via above results. Here, we used linear mixed effects model in order to consider the effect of individual by adjustment of random effect. In terms of relative quality and prediction, our proposed method outperformed the models without clustering information according to AIC standards and prediction error.