Development of 365-day Life Expectancy Models: Application of Machine Learning Methods to a Prospective Study of Critically Ill Cirrhotic Patients (Preprint)

Abstract

The mortality rate of cirrhotic patients in intensive care units (ICUs) is usually high. Analyzing high dimensional data and incorporating accurate life expectancy indices using machine learning methods may improve the accuracy of the prediction of the long-term prognosis of critically ill cirrhotic patients.

To develop highly accurate 365-day candidate life expectancy models based on machine learning methods, and compare their accuracy with CLIF-SOFA scores in critically ill cirrhotic patients.

This study analyzed 141 prospective predictor variables on day 1 of 294 cirrhotic patients admitted to a 10-bed specialized hepato-gastroenterology ICU in a 2000-bed tertiary care referral hospital from September 2010 to August 2013. The least absolute shrinkage and selection operator (LASSO) method was used to select subsets of predictor variables. Model validation was applied to compare multiple life-expectancy models with CLIF-SOFA scores.

The overall in-hospital and 365-day mortality rates were 55.4% (163/294) and 82.3% (242/294), respectively. CLIF-SOFA score, hepatic encephalopathy grade and serum creatinine were the most critical risk predictors. The final models were developed by incorporating the most important 16 predictors based on Support Vector Machine (SVM), Random Sample Consensus (RANSAC), Ensemble, Random Forest and Decision Tree methods. The SVM model showed excellent discrimination (0.901±0.021) and outperformed CLIF-SOFA score (0.841±0.024) as well as the others. The RANSAC model had the best Youden index, and the Ensemble model had the highest overall correctness of prediction.

The 365-day life expectancy models based on SVM/Random Forest/RANSAC/Ensemble classifiers were excellent prognostic evaluation tools for critically ill cirrhotic patients.