Clinically meaningful risk factors for recurrence in T1 colorectal cancer treated with endoscopic resection alone identified by unsupervised machine learning: a multicenter study

Background and study aim: Identifying high-risk patients for recurrence after endoscopic resection (ER) of T1 colorectal cancer (CRC) remains challenging. This study aimed to identify recurrence-risk subtypes and develop an interpretable risk stratification framework. Patients and methods: This retrospective study analyzed 1123 patients with T1 CRC treated with ER alone across 27 Japanese institutions (July 2009-December 2016). Patients were divided into development (68%) and evaluation (32%) cohorts based on institutional stratification. K-means clustering was applied to clinicopathological variables to identify recurrence-risk subtypes. A decision tree classifier was subsequently developed to generate transparent risk stratification rules. Results: Three distinct subtypes were identified in the development cohort. Subtype 1 exhibited a numerically higher recurrence rate (5.4%) than subtype 2 (0.9%) and subtype 3 (1.4%). Although subtypes 2 and 3 showed comparable recurrence rates, they were clearly differentiated by morphology (flat vs. polypoid). In the evaluation cohort, subtype 1 continued to show a numerically higher recurrence (4.8%) compared with subtypes 2 (1.6%) and 3 (1.1%). The decision tree model stratified recurrence risk hierarchically: submucosal invasion <1,000μm indicated low risk, whereas invasion ≥1,000μm required morphological assessment, with polypoid lesions classified as high risk and flat lesions further stratified using a 2,000μm threshold. Conclusions: Three clinically distinct recurrence risk subtypes were identified in T1 CRC following ER, suggesting that morphological subclassification of T1b lesions may refine stratification beyond conventional depth-based criteria. The decision framework offers a preliminary exploratory basis for recurrence risk assessment in this population.