Area Deprivation and Patient Complexity Predict Low-Value Healthcare Utilization in Persons with Heart Failure

Background Heart failure (HF) is a debilitating condition affecting over 6.7 million adults in the United States. Social risks and complexity, or personal, social, and clinical aspects of persons' experiences, have been found to influence healthcare utilization and hospitalizations in persons with HF. Low-value utilization, or irregular outpatient visits with frequent emergency room use, or hospitalization is common among persons with complex conditions and social risk and requires further investigation in the HF population. Objectives The purpose of this research was to assess the influence of complexity and social risk on low-value utilization in persons with HF using machine learning approaches. Methods Supervised machine learning, tree-based predictive modeling was conducted on an existing data set of adults with HF in the eight-county region of Western New York for the year 2022. Decision tree and random forest models were validated using a 70/30 training/testing data set and k-fold cross-validation. The models were compared for accuracy and interpretability using the area under the curve, Matthews correlation coefficient, sensitivity, specificity, precision, and negative predictive value. Results Area deprivation index, a proxy for social risk, number of chronic conditions, age, and substance use disorders were predictors of low-value utilization in both the decision tree and random forest models. The decision tree model performed moderately, whereas the random forest model performed excellently and added hardship as an additional important variable. Discussion This is the first known study to look at the outcome of low-value utilization, targeting individuals who are underutilizing outpatient services. The random forest model performed better than the decision tree; however, features were similar in both models, with area deprivation index as the key variable in predicting low-value utilization. The decision tree was able to produce specific cutoff points, making it more interpretable and useful for clinical application. Both models can be used to create clinical tools for identifying and targeting individuals for intervention and follow-up.