Machine Learning Holds Promise for Predicting Acute Care Risk

Presenting author: Marianna Elia, MS

Approximately 10%-20% of patients undergoing outpatient radiotherapy or chemoradiation require acute care (emergency department evaluation or hospitalization) due to treatment-related toxicities, disease progression or comorbidities. Reducing acute care events is clinically and economically important, as it may improve outcomes, enhance quality of life, align with patient preferences, and decrease costs for both patients and the health care system.

Machine learning (ML) has emerged as a powerful tool for predicting emergency department (ED) visits and other acute care events across diverse clinical domains, including oncology, chronic disease management and perioperative care. By integrating large, heterogeneous datasets — such as demographics, comorbidities, vital signs, laboratory results and medical history — ML models can identify complex, non-linear risk patterns often missed by traditional statistical methods. Many models achieve strong predictive performance, with AUROCs typically ranging from 0.75 to 0.90.

Although most studies remain retrospective, SHIELD-RT represents one of the first prospective, interventional ML trials, providing valuable insights into clinical workflow integration and resource allocation.

In this multi-institutional evaluation, the SHIELD-RT model was tested retrospectively across two more independent cancer centers with distinct patient populations, clinical workflows and EHR systems. Analyzing more than 22,000 radiotherapy courses, the model achieved AUROCs of 0.756 and 0.770, demonstrating good calibration and clear risk discrimination. These results mirrored the original prospective randomized trial, which reduced acute care events by 45% and costs by 48%. Collectively, these findings underscore the model’s generalizability and highlight its potential for broad clinical implementation.

“It takes us one step closer to a routine clinical integration of the SHIELD-RT model to improve care at any hospital,” shared Elia. “The validation phase, which in this case is retrospective, is an often overlooked but critical aspect of ensuring that a ML model will work across different clinical settings — second only to choosing the right problem to solve and equipping the model with the right data to solve it. The model was prospectively validated during the SHIELD-RT randomized control trial at Duke, where the model was trained, and now the focus has shifted to externally validate the model at different hospitals retrospectively.”

In summary, this study, presented by Marianna Elia, MS, on Monday, September 29 at ASTRO 2025, delivers strong confirmation that the SHIELD-RT machine learning model can reliably identify radiotherapy patients at high risk of acute care events across diverse health care settings. Its adoption could improve patient outcomes while reducing costs, aligning with the growing emphasis on value-based care. As health care systems increasingly transition toward population health strategies, ML-driven risk prediction is poised to become an integral component of modern oncology practice.

Abstract 180, Multi-Institutional Validation of the SHIELD-RT Machine Learning Model to Prevent Acute Care Events during Radiotherapy, was presented during the SS 13 - DHI 1: The Digital Revolution in Radiation Oncology: AI Models for Enhanced Patient Care session of the 67th ASTRO Annual Meeting.