Author: Shahed Badiyan, Bin Cai, Tu Dan, Michael Dohopolski, Steve B. Jiang, Deepkumar Mistry, Arnold Pompos, Robert Timmerman, Jing Wang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab & Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, University of Texas Southwestern Medical Center, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Biology-guided radiotherapy (BGRT) offers significant potential for personalized and adaptive cancer treatment, with clinically available systems such as SCINTIX from Reflexion now being introduced into clinical practice. However, BGRT systems have specific patient eligibility criteria—for instance, SCINTIX requires a minimum standardized uptake value (SUV) of 6 on diagnostic PET. Manually identifying BGRT-eligible patients can be challenging in a busy clinical setting, potentially causing some patients to miss the opportunity for advanced treatment. To address this, we aim to develop an autonomous AI agent that extracts radiology reports and alerts clinicians to potential BGRT-eligible patients before their consultation, streamlining patient selection and improving access to BGRT.
Methods: The AI agent first compiles a list of patients scheduled for consultation the following day from the scheduling system (Mosaiq). It then extracts radiology imaging narratives from Epic and identifies patients with SUV > 6. To further refine eligibility, an OpenAI large language model (LLM), implemented in the Azure environment, is used to identify patients with lung and bone cancer, key criteria for SCINTIX treatment. The system then consolidates the identified patients' characteristics along with relevant portions of the radiology narrative related to SUVs and sends this information via email to physicians.
Results: This automated AI agent and notification system has been implemented in our clinic. Among 6,305 patient records reviewed between August 2024 and January 2025, 1,884 were identified with SUV > 6. Using Azure OpenAI LLM, we have further identified 241 Bone and 379 Lung patients. The key information summarized in the email enables physicians to quickly screen potential BGRT candidates.
Conclusion: The system successfully streamlined the identification and prioritization of patients eligible for the BGRT. By automating this process, physicians receive timely and accurate information about high-priority cases, improving clinical workflows and ensuring patients receive advanced treatment.