Validation of DiaSurv, an AI-based algorithm for stage II colon cancer risk stratification.

Abstract

3524 Background: Adjuvant chemotherapy decisions in stage II colon cancer (IIA T3N0; IIB T4aN0; IIC T4bN0) remain challenging, as conventional risk factors poorly predict recurrence. This results in both overtreatment and undertreatment. AI-based analysis of routine H&E slides may enable more accurate risk stratification to guide personalized treatment decisions. Methods: DiaSurv Colon is an AI-powered prognostic tool using a deep neural network to extract in an unsupervised way morphological features from H&E-stained WSI and predict 5-year overall survival (OS). Each patient is assigned an individual survival risk score. The model was trained on the TCGA cohort (n = 463) and previously validated on two external cohorts. In this study, we evaluated its performance on two new independent external cohorts, digitized with two different scanners: CHU Caen (n = 201; 5-y OS 47%, 95%CI 39-53%) and IHP (n = 207; 5-y OS 54%, 95%CI 45-61%). Prognostic accuracy was assessed using concordance index (c-index). Multivariate Cox regression and log-rank tests with hazard ratios (HR) were used to compare DiaSurv Colon performance against conventional clinicopathological factors. Results: On external validation, DiaSurv Colon achieved a c-index of 58 (CHU Caen) and 64 (IHP). The AI-based risk score significantly stratified patients: 5-year OS was 64% (95%CI 50-75%) vs 40% (95%CI 32-48%) for low- vs high-risk groups in CHU Caen (p < 0.01), and 64% (95%CI 52-73%) vs 39% (95%CI 26-51%) in IHP (p < 0.001). The risk score demonstrated strong prognostic value with hazard ratios of 1.29 (95%CI 1.05-1.58, p < 0.01) and 1.57 (95%CI 1.23-2.00, p < 0.001) for CHU Caen and IHP, respectively. Conclusions: DiaSurv Colon demonstrated consistent prognostic performance across independent external cohorts, with robust hazard ratios and concordance indices. This AI-based tool may improve risk stratification in stage II colon cancer and support personalized adjuvant therapy decisions with more confidence.