A causal multi‑modal AI model leverages routinely collected pathology and clinical data to predict personalized chemosensitivity. The model was trained on a multinational cohort of 9,141 patients (12 cohorts across nine countries) and evaluated on an independent set of 1,994 patients (5 cohorts in three countries). It outputs treatment‑specific recurrence probabilities for each individual, achieving near‑perfect calibration and strong prognostic discrimination over both 5‑year and 10‑year follow‑up horizons.
Chemotherapy benefit predictions exhibit robust performance and consistently outperform existing recurrence‑score‑based tests. Compared with standard care, using the model to guide personalized decisions could reduce the proportion of patients receiving chemotherapy by roughly 30% while maintaining the same recurrence‑free survival rate. Tumors identified as highly chemosensitive share molecular and morphological programs of proliferation, cell‑cycle progression, and replication stress.
Importantly, the model transfers zero‑shot to non‑breast cancers, suggesting that the causal multi‑modal AI approach may serve as a universal strategy for predicting treatment outcomes across cancer types.
Review