Leveraging AI to enhance symptom capture and reduced hospitalizations.

1572 Background: Unplanned hospitalizations and ED visits for cancer patients impose significant morbidity, financial costs, and reduced quality-of-life. Efficient resource allocation in oncology care requires proactive strategies to identify patients at high risk for preventable admissions, optimizing bed availability and outpatient management. CMS classifies acute care utilization (ACU), hospital stays or ED visits, within 30 days of chemotherapy for certain conditions, as “preventable” under OP-35, emphasizing the need for better risk stratification. By leveraging artificial intelligence (AI) including machine learning and large language models (LLM), predictive models can analyze the electronic health record (EHR) to identify patients who may benefit from early outpatient interventions. This study leverages AI to assess preventable admissions and quantify the benefits of proactive management, enhancing patient outcomes and care efficiency in oncology. Methods: This study analyzed data from 18,187 patients across a multisite cancer center to develop predictive models for ACU within 30 days of any systemic therapy following OP-35 criteria. Models incorporated structured data and clinical notes using LLMs. Using 2010–2019 data for training and 2020–2024 for validation, XGBoost and Random Forest models were developed to maximize sensitivity while maintaining acceptable specificity. Estimates of preventable hospital bed days were modeled to evaluate the impact of AI-driven risk stratification and targeted interventions on hospital resource utilization. Results: The 30-day hospital visit prediction model demonstrated strong performance, with the XGBoost algorithm achieving an AUROC of 0.84 (95% CI: 0.83–0.86). Incorporating later therapy lines improved accuracy by accounting for the complexity of advanced disease. A threshold was set to prioritize sensitivity for identifying high-risk patients while maintaining specificity to minimize unnecessary interventions. Model implementation was estimated to prevent 22% of hospital visits when paired with timely intervention, saving 1,160 of the 5,370 bed days observed in the 2021–2024 cohort. Conclusions: This study underscores the potential of AI-driven prediction models to enhance precision oncology by identifying patients at risk for unplanned hospital visits following systemic cancer therapy—a critical quality indicator impacting patient outcomes, healthcare costs, and operational efficiency. By incorporating multi-line therapy data and leveraging advanced modeling techniques, the approach effectively captures disease progression and personalized treatment histories. Utilizing LLMs to structure fragmented data across care systems addresses a prevalent challenge in oncology. Accurate risk prediction of hospitalization facilitates proactive interventions, improves care coordination, reduces bed occupancy, and supports informed decision-making, ensuring timely and targeted support for high-risk patients.