Deep Learning-Based Multi-Cancer Analysis for Predicting Disease-Free Survival Across Multiple Cancer Types

Abstract

Simple Summary Artificial intelligence-derived parameters hold promise for tumor prognosis, yet their application across cancer types remains underexplored. We developed a deep learning-based multi-cancer disease-free survival (MC-DFS) model using 8856 cases with whole-slide images and clinical data. The training cohort comprised 7392 cases from TCGA (24 cancer types), and independent external validation included 1464 cases from CPTAC and a hospital cohort (9 cancer types). Subsequently, a nomogram (NOMO) integrating MC-DFS, tumor stage, and age was constructed. In the training and validation cohorts, MC-DFS achieved AUCs of 0.750 and 0.682, respectively, with hazard ratios of 4.823 and 2.092 (both p < 0.0001). The model performed robustly across cancer subtypes, with the nomogram improving risk stratification, particularly for stage I malignancies, complementing existing staging. With multicenter validation, this system could become a practical tool for managing diverse cancers.