Abstract 7428: Predicting acute care utilization following cancer treatment using machine learning: A patient-centered approach to precision oncology

Abstract Introduction: Unplanned acute care utilization (ACU) after receipt of cancer treatment is a significant challenge in oncology, often reflecting care gaps and driving higher costs. To address this, the Centers for Medicare & Medicaid Services (CMS) have designated ACU as a critical quality metric (OP-35). Predicting patients at risk and understanding root causes of ACU can enable interventions that improve outcomes for patients. This work applies advanced machine learning models, integrating structured and unstructured data, to predict ACU in diverse patients across varying malignancy categorizations and treatments. Methods: This study utilized data on 18, 143 patients from a comprehensive, multisite cancer center, to develop machine learning models predicting ACU within 30 days of cancer therapy initiation per OP-35 criteria. Predictions for each therapy line were generated separately. Predictions utilized structured data from medical records and when earlier therapy lines were received outside the center, large language models were applied to clinical notes to encode this unstructured data into the model. Demographic variables, including race, ethnicity, and insurance, were included to identify potential biases. The model, trained on 2010-2019 data and validated on 2020-2024 data, used mean imputation for missing values and employed XgBoost and Random Forest algorithms for supervised learning with hyperparameter tuning to ensure the best trade-off between bias and variance in the model. Results: The 30-day ACU prediction model showed strong predictive performance, with an XGBoost model achieving an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.84 (95% CI: .82-.85). Performance was further analyzed across OP-35 categories, with the highest accuracy in predicting ACU due to anemia (AUROC: .87, 95% CI: .86-.88) and infection (AUROC: .86, 95% CI: .85-.87). Incorporating later therapy lines enhanced predictive accuracy by capturing the complexities of advanced disease. Conclusions: This study highlights the potential of ML to advance precision oncology by predicting ACU following cancer therapy—a critical indicator of patient-centered care quality. By incorporating data from multiple therapy lines and leveraging advanced modeling techniques, it addresses complexities such as disease progression and diverse treatment histories. Decomposing predictions by ACU causes offers actionable insights for timely interventions. These findings highlight the importance of integrating advanced analytics into clinical workflows to improve outcomes, reduce costs, and support data-driven decision-making in oncology. Citation Format: Arman Koul, Mohana Roy, Behzad Naderalvojoud, Tina Hernandez-Boussard. Predicting acute care utilization following cancer treatment using machine learning: A patient-centered approach to precision oncology [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7428.