A Probabilistic Machine Learning Framework for Early Disease Prediction and Clinical Risk Stratification Using Statistical Inference and Artificial Intelligence

Abstract

This paper introduces a probabilistic machine-learning approach to nonlinear minimax optimization, statistical inference, and clinical risk stratification of early adverse clinical outcome prediction. Diabetes 130-US Hospitals dataset containing 101,764 encounters of 71,518 patients was used to assess the framework to predict 30-day hospital readmission. A neural-network model minimizing the smooth-max using the minimax-optimization was compared to logistic regression, random forest, gradient boosting, and Bayesian logistic regression. The assessment of performance was based on the use of the following measures of performance: AUROC, AUPRC, Brier score, worst-group loss, bootstrap inference, and sensitivity analysis of age-distribution. The proposed model achieved an AUROC of 0.6380, AUPRC of 0.1850, Brier score of 0.0927, and worst-group loss of 0.3441. Gradient boosting had a slightly higher discrimination (AUROC = 0.6416), with no statistically significant higher discrimination proven by the proposed model. Clinical risk stratification was used to define low-, moderate-, and high-risk groups with reported readmission rates of 5.76%, 10.40%, and 15.96% respectively. The results show that subgroup-oriented minimax estimation is feasible to predict probabilistic clinical risks with a focus on the necessity of additional external and clinical validation.