Identifying Post-Surgical Recurrence Subtype of T1-Stage Colorectal Cancer by Machine Learning

INTRODUCTION: Traditional risk stratification heavily relies on expert judgment and manually established thresholds. This study aims to automatically identify subtypes in the patients of T1-stage colorectal cancer with distinct clinicopathologic characteristics and recurrence risk profiles, using machine learning. METHODS: We analyzed data from 3,367 patients (mean follow-up, 1,281 days) with T1 colorectal cancer who underwent surgical resection between 2009 and 2016 across 27 high-volume core Japanese institutions. Patients were split into derivation and test datasets (4:1 ratio). Hierarchical clustering was employed to identify recurrence subtypes in the derivation dataset. Machine learning classifiers were developed and validated on the test dataset. Co-occurrence and Bayesian network analyses aided interpretation. RESULTS: Three distinct subtypes were identified: two high-risk (subtypes 1 and 2) and one low-risk (subtype 3). Subtype 1 was predominantly associated with polypoid morphology (94.8%), whereas subtype 2 was characterized by flat morphology (89.4%). Subtype 2 showed a relatively consistent presence across most factors, with comparable levels of lymphatic invasion, vascular invasion, and tumor budding. Subtype 3 shared similarities with subtype 1 in polypoid morphology (76.5%) but differed in other factors. These findings showed similar trend on the test dataset. Subtype-specific risk factors included lymphovascular invasion and nodal metastasis in both high-risk subtypes, while rectal location was unique to subtype 1 and polypoid morphology and large size were specific to subtype 2. CONCLUSION: This machine learning approach identified three distinct recurrence subtypes of T1 colorectal cancer, each with unique characteristics and risk profiles, indicating the potential value of subtype-specific clinical strategies.