Can Cluster-Boosted Regression Improve Prediction of Death and Length of Stay in the ICU?

Sharing of personal health information is subject tomultiple constraints, which may dissuade some organizations from sharing their data. Summarized deidentified data, such as that derived from k-means cluster analysis, is subject to far fewer privacy-related constraints. In this paper, we examine the extent to which analysis of clustered patient types can match predictions made by analyzing the entire dataset at once. After reviewing relevant literature, and explaining how data are summarized in each cluster of similar patients, we compare the results of predicting death, and length of stay (LOS) in the ICU1 using regression analysis on original and clustered data from the MIMIC II dataset. Clustering improved regression prediction accuracy for both death and LOS. We then show that clustering prior to regression also improved prediction of number of days to next emergency room visit for cancer patients. Thus, in all three prediction tasks that we investigated (involving two very different datasets), we found that clustering prior to regression analysis improved prediction accuracy. We discuss the results in terms of their implications for the future use of health-repository-based data analytics to provide a supplement to existing methods of clinical decision support.