A generative model–assisted clinical decision support system for the postoperative management of colon cancer.

Abstract

e15508 Background: Postoperative management of colon cancer relies on the accurate interpretation of complex pathology reports and adherence to evolving clinical guidelines. However, manual assessment is time-consuming and prone to numerical errors, resulting in staging misclassification and guideline-discordant therapy. Utilizing Large Language Models (LLMs) to bridge the gap between raw clinical data and standardized decision-making, while enhancing patient communication, remains an unmet clinical need. Methods: We developed and verified an LLM-based Clinical Decision Support System (CDSS) utilizing a dataset of 4,608 pathology reports from two tertiary hospitals and 409 cases from TCGA. The system implements a standardized "pathology-to-decision" workflow comprising five key modules: (1) Automated Staging: Extraction of unstructured data to automate TNM staging (AJCC 8 th edition) with explicit reasoning and field tracing; (2) Risk Stratification: Classification of MMR status and assessment of high-risk factors (e.g., lymph nodes < 12, perineural invasion) to distinguish high-risk from low-risk patients; (3) Decision Support: Generation of guideline-concordant treatment recommendations and follow-up schedules calculated from the surgery date (NCCN Guidelines 2025 v4); (4) Prognostic RAG: A PubMed-based Retrieval-Augmented Generation module that synthesizes patient demographics and pathology features to answer queries regarding 5-year survival, recurrence windows, and genetic screening necessity; and (5) Patient Education: Simplification of complex reports into patient-friendly summaries. Results: Comparative analysis with manual expert review demonstrated that the CDSS significantly reduced the average time required for case processing by 98.3% (approx. 9.8s vs. 578.9s per case). In performance validation, the system achieved an accuracy of 96.8% for TNM staging and 97.9% for risk stratification, and reduced guideline-discordant treatment decisions. The PubMed-RAG module provided evidence-based prognostic information, and the patient summaries showed improved readability scores compared to original reports (Table 1). Conclusions: The CDSS improves postoperative staging accuracy, supports guideline-concordant treatment and surveillance, and facilitates patient-centered risk counseling. It provides a scalable approach to standardizing management of colon cancer across diverse clinical settings. Comparison of postoperative decision support approaches. Feature CDSS Manual Practice Staging Logic Deterministic (Rule-based) Experience-dependent Guideline adherence Real-time (AJCC/NCCN) Manual updates Traceability Fully traceable (Linked to source) Manual cross-checking Multilingual pathology support Yes Limited Patient support Integrated, guideline-based Time-consuming Time per Case < 10 seconds 5-10 minutes