An Artificial Intelligence (AI) model integrating multiscale foundation model histopathology representations with molecular and clinical features predicts early and late distant recurrence in TAILORx

We developed and validated IICM+, a multimodal model integrating clinicopathologic variables, transcriptomic features, and multiscale histopathology-derived image representations to predict distant recurrence in hormone receptor-positive, HER2-negative, node-negative early breast cancer. Image features included tile-level embeddings and slide-level representations generated by a custom multimodal foundation model pretrained on paired histopathology images, RNA sequencing, and pathology reports. Using TAILORx specimens with long-term follow-up, models were trained in a development cohort with five-fold cross-validation ( n = 2808) and evaluated in an independent institutional holdout validation set ( n = 1621). In holdout validation, IICM+ showed strong prognostic discrimination for overall distant recurrence (C-index 0.735, 95% CI 0.681–0.782), early distant recurrence (0.791, 95% CI 0.715–0.858), and late distant recurrence (0.710, 95% CI 0.645–0.773). IICM+ separated high- versus low-risk groups for overall distant recurrence (HR 5.25, 95% CI 3.50–7.86; P < 0.001) and remained prognostic (HR 3.56, 95% CI 2.22–5.70, P < 0.001) after adjustment for clinicopathologic covariates and RS category. IICM+ also identified RS/IICM+ discordant groups with different observed recurrence risks, supporting additional prognostic stratification beyond RS.