Statistical Approaches to Evaluating Artificial Intelligence Models in Healthcare

Abstract

This study reviews statistical frameworks utilized in an evaluation of artificial intelligence (AI) modeling ability to identify death in heart failure patients. Descriptive statistics, and correlation analyses had been employed to describe demographic and clinical variables (age, ejection fraction, serum creatinine, and comorbidity) that depict the sample, through analyzing a sample size of 299 observations. The distribution of the outcome suggested a 32% death rate indicative of the severity of condition. Logistic Regression and Random Forest model were used to assess predictive performance. Logistic regression produced a predictive accuracy of 0.83 with AUC of 0.86 while Random Forest had a predictive accuracy of 0.81 AUC of 0.89. The Precision-Recall curve analysis was suggestive of a stronger predictive classification irrespective of class imbalance. The results of this research indicate traditional statistical modeling can be coupled with machine learning to enhance predictive reliability and clinical utility. This study also highlighted the advantages of interpretive data driven approaches for health analytics.