Mode of presentation and symptoms of colorectal cancer (CRC) by primary tumor and molecular subgroup: A large-scale AI-powered analysis.

Abstract

203 Background: The mode of presentation of CRC has implications for early detection and prognosis. Our aim was to characterize presentation and symptoms at CRC diagnosis by primary location, RAS / BRAF , and mismatch repair (MMR) status at scale. Methods: A retrospective cohort of CRC patients (pts) diagnosed 1990-2025 was collected. Presentation mode (symptomatic, emergency, screening, incidental, or unclear) and symptoms were extracted using the GPT-4o large language model (LLM) with prompt engineering on manually curated clinical notes relevant to diagnosis. Clinical and molecular data were obtained from institutional datasets. Chi-square and relative risk were used for comparisons. Results: In 41,152 pts, LLM-based classification showed a high concordance with manual review (Cohen’s κ = 0.90). Emergency presentation was more common in right and left colon than rectum (28% vs 23% vs 9%, p < 0.01). Conversely, symptomatic presentation was more frequent in rectum than left and right colon (81% vs 66% vs 57%, p < 0.01). Rectal cancers most commonly presented with bleeding-related symptoms and changes in bowel habits, while abdominal/rectal pain, constitutional symptoms, and anemia were most frequent in right colon primaries (Table). In 19,321 pts with MMR status, dMMR were twice as likely to present with anemia and 37% more likely with constitutional symptoms ( p < 0.01), whereas pMMR had a 44% and 33% higher risk to present with bleeding-related symptoms and changes in bowel habits ( p < 0.01). In 9,525 pts with RAS / BRAF status, BRAF -V600E more often presented emergently compared with RAS mt and RAS & BRAF wt (31% vs 26% vs 23%, p < 0.01), while symptomatic presentation was enriched in RAS & BRAF wt and RAS mt compared with BRAF -V600E (66% vs 63% vs 57%, p < 0.01). Bleeding-related symptoms and changes in bowel habits were most common in RAS & BRAF wt, while abdominal/rectal pain and anemia were more frequent in RAS mt and BRAF -V600E (Table). No meaningful differences were seen in screening or incidental cases. Conclusions: LLMs enable reliable large-scale oncology research. Mode of presentation and symptoms of CRC differ substantially by primary site and molecular profile, highlighting the connection between biological and anatomical determinants of CRC diagnosis. Right colon Left colon Rectum p RAS & BRAF wt RAS mt BRAF -V600E p Total n 7478 9928 7715 2967 3413 325 Bleeding-related* 1248, 17 4626, 47 5842, 76 <0.01 1304, 44 1317, 39 73, 22 <0.01 Abdominal/rectal pain 3658, 49 4037, 41 1762, 23 <0.01 1264, 43 1551, 45 171, 53 <0.01 Changes in bowel habits 1072, 14 3067, 31 3065, 40 <0.01 936, 32 905, 27 59, 18 <0.01 Constitutional 1515, 20 1403, 14 832, 11 <0.01 467, 16 550, 16 55, 17 0.82 Anemia 2238, 30 850, 9 432, 6 <0.01 322, 11 553, 16 66, 20 <0.01 Obstruction 1127, 15 1573, 16 291, 4 <0.01 432, 15 480, 14 51, 16 0.67 Nausea/vomiting 838, 11 662, 7 130, 2 <0.01 189, 6 295, 9 30, 9 <0.01 For symptomatic/emergency. *Data shown as n, %.