Abstract
BACKGROUND: Rising healthcare costs worldwide have spurred interest in using machine learning to pinpoint major healthcare utilizers for population health efforts. Previous studies have focused on individuals who impose the highest financial burden, such as targeting high-need high-cost utilizers, a small group with limited potential for cost reduction. Therefore, we developed models to predict future healthcare utilization across various thresholds. METHODS: We leveraged data from a multi-institutional diabetes database to develop binary classification models capable of predicting healthcare utilization in the following year. The models predicted six different outcomes: patients with an annual total length of stay of ≥ 7, ≥14, and ≥30 days, and emergency department (ED) attendance of ≥ 3, ≥5, and ≥10 visits. To mitigate class imbalance, we applied random and synthetic minority oversampling techniques and compared against models trained without oversampling. Models were trained with 2019 data and tested on unseen data from 2020 and 2021. RESULTS: Models trained with random oversampling, including logistic regression, multivariate adaptive regression splines, boosted trees, and multilayer perceptron, consistently demonstrated high AUC (>0.80) and sensitivity (>0.60) across both training-validation and test datasets. Addressing class imbalance was found to be crucial, as models without oversampling displayed satisfactory AUC (>0.80) but significantly lower sensitivity (<0.10). Key predictors identified included age, number of ED visits in the present year, chronic kidney disease stage, inpatient bed days in the present year, and mean HbA1c levels. CONCLUSIONS: Our machine learning models successfully predict high service level utilization with strong performance, offering valuable tools for policymakers and health planners to develop targeted health initiatives based on patient utilization patterns. KEY MESSAGES: • Machine learning models can be leveraged to predict future healthcare utilization and facilitate interventional program development. • Annual total length of stay and total emergency department visits are useful service level indicators.