S1414
Interdisciplinary - Health economics & health services research
ESTRO 2026
delivery with strategic goals. Our experience underscores the importance of early IT involvement, robust governance structures, and regional data- sharing agreements to ensure sustainability and scalability. When these conditions are met, VBHC has the potential to optimize outcomes and reduce costs. Keywords: VBHC, patient-relevant outcomes, PDSA- cycle Proffered Paper 3089 Large Language Model-Assisted CTCAE Toxicity Scoring from Clinical Conversations Federico Mastroleo 1,2 , Mariana Borras-Osorio 1 , Mi Zhou 1 , Satomi Shiraishi 1 , Andrew Y.K. Foong 1 , David Routman 1 , Mark R. Waddle 1 1 Department of Radiation Oncology, Mayo Clinic, Rochester, USA. 2 Division of Radiation Oncology, IEO, European Institute of Oncology, IRCCS, Milan, Italy Purpose/Objective: Accurate toxicity assessment is essential for safe and effective radiotherapy delivery, yet current documentation is often incomplete and inconsistent due to inter-observer variability and manual abstraction. Existing Natural Language Processing and Large Language Models (LLM)-based approaches have shown potential for extracting adverse events from electronic records but remain limited by dependence on structured text and lack of real-time applicability. Conversational recordings between clinicians and patients represent an untapped source of rich, temporal information on treatment-related toxicities. This pilot study aimed to develop and evaluate an automated pipeline leveraging speech recognition and LLMs to extract CTCAE toxicities directly from clinical encounter recordings, enabling scalable and standardized toxicity documentation within routine workflows. Material/Methods: Audio recordings of physician/nurse and patient interactions were securely collected and stored in a compliant environment. Each recording was transcribed using the OpenAI Whisper Large-v3 model to generate accurate text outputs, followed by speaker diarization with WhisperX, to distinguish between patient and clinician utterances. The processed transcripts were analyzed using Gemini 2.0 Flash for context-aware speaker recognition and clinical dialogue understanding. Subsequent Q&A extraction and data grounding modules identified exchanges referring to symptoms, adverse events, or toxicity descriptions. The refined outputs were processed with Gemini 2.5 Pro for toxicity extraction, mapping identified terms to CTCAE categories and grades. A data reconciliation stage consolidated outputs across
adverse events, measured routinely during the patient journey and visualized through interactive dashboards. A structured Plan-Do-Study-Act cycle guided performance analysis and innovation (Figure 1). The implementation starts with the visualization of patient-relevant outcomes and a detailed analysis of our institute’s performance, benchmarked against national and international standards. These insights inform the formulation of targeted improvement goals and the selection of evidence-based innovations. Regular evaluations of the impact on outcomes allow for timely adjustments. IT infrastructure plays a critical role in data visualization and feedback loops. Facilitators and barriers were thoroughly identified based on our experience with implementation of the methodology for prostate cancer through weekly observations and in-depth team discussions with clinicians and IT staff.
Results: A real-time dashboard for prostate cancer was developed. The dashboard reveals whether our care quality is consistent with outcome levels observed in national and international benchmarks. It enables stratified performance analysis of the patient-relevant outcomes based on patient characteristics and treatment pathways.Key facilitators of our methodology for implementing VBHC included dedicated IT capacity and project managers to facilitate implementation and strong clinical leadership. Visualization of outcomes in dashboards fostered engagement among clinicians. However, several barriers were identified in the usability, completeness and reliability of clinical data for steering towards quality improvement: a) clinical data are recorded for individual care, not population-level analysis; b) some outcomes required complex data transformations across time points; c) variations in PROMs data collection and major IT transitions required meticulous harmonization of data; d) data sharing between referring hospitals and our institute proved challenging due to privacy regulations. Conclusion: VBHC implementation in radiotherapy is both feasible and promising. It enhances transparency, supports clinician-led quality improvement, and aligns care
Made with FlippingBook - Share PDF online