Machine Learning for Predicting Readmission Risk Among the Frail - Explainable AI for Healthcare

Abstract

Healthcare costs that can be attributed to unplanned readmissions are staggeringly high and negatively impact health and wellness of patients. In the United States, hospital systems and care providers have strong financial motivations to reduce readmissions in accordance with several government guidelines. One of the critical steps to reducing readmissions is to recognize the factors that lead to readmission and correspondingly identify at-risk patients based on these factors. While hospital readmission is an undesirable outcome for any patient, it is more so for medically frail patients. Here, we present readmission risk prediction using five machine learning approaches - Random Forest, XGBoost, CatBoost, and Logistic Regression, and a Stacking classifier for predicting 30-day unplanned readmission for patients deemed frail (Age ≥= 50). We use a comprehensive and curated set of variables that include frailty, comorbidities, high risk medications, demographics, hospital, and insurance utilization to build these models. Findings indicate that the CatBoost model outperforms other models with a mean AUC of $0.79$. We find that prior readmissions, discharge to a rehabilitation facility, length of stay, comorbidities, and frailty indicators were all strong predictors of 30 day readmission. Since explainability is paramount for healthcare, we present in depth insights into the model's predictions using SHapley Additive exPlanations (SHAP) — the state of the art in machine learning explainability. We highlight important features for prediction, showcase the distributions of those features in the data and how feature observations impacted the risk of readmission. We select and observe the relative variable influence of patient observations with the highest and lowest readmission risk, and show how the model made predictions in order to demonstrate explainability at an individual observation level. Finally, we identify six patient groups with different levels of readmission risk using clustering techniques.