Abstract
Rising healthcare costs and a shortage of primary care providers in the United States create substantial strain on the healthcare system, underscoring the need for efficient allocation of limited resources. Accurate prediction of high primary care utilization can enable proactive care planning, targeted interventions, and workload optimization. We developed and evaluated the Friedman Score, a machine learning-based model that predicts estimated yearly primary care utilization categories, low use (0-85th percentile), High Use (>85-95th percentile), and very High Use (>95th percentile), using structured electronic health record data from UCSD Health primary care patients in 2022-2023. Features included age, chronic disease diagnoses, medication history, and acute care patterns. XGBoost was selected as the primary modeling approach, and its results were benchmarked against five other machine learning algorithms. Across both years, XGBoost consistently demonstrated high discriminative ability (AUC 0.78-0.89 in 2022; 0.81-0.89 in 2023) and robust calibration. SHAP analysis identified medication usage, age, and chronic disease burden, particularly depression, as the most influential predictors. The Friedman Score offers a robust, interpretable tool for identifying high-utilization patients, providing actionable insights to guide proactive, data-driven primary care delivery.