Causal Multimodal AI Predicts Personalized Chemotherapy Sensitivity

Summary

Chemotherapy can improve survival for some breast cancer patients, but clinicians cannot reliably determine who will benefit. Researchers present a causal multimodal AI model that uses routinely collected pathology and clinical information to estimate treatment-specific recurrence probabilities and personalized chemosensitivity. The model was developed using data from 9,141 patients across 12 cohorts in nine countries and evaluated on a separate set of 1,994 patients from five cohorts in three countries. It showed near-perfect calibration and strong prognostic discrimination at both five- and ten-year follow-up horizons. Its predictions of chemotherapy benefit also performed robustly and outperformed existing recurrence-score-based tests. The authors report that using the model to support treatment decisions could reduce chemotherapy use by 30% while maintaining the same recurrence-free rate as standard care. Tumors classified as highly chemosensitive had molecular and morphological features associated with proliferation, cell-cycle progression, and replication stress. The model also transferred zero-shot to non-breast cancers, which the authors say suggests a potentially general strategy for predicting treatment outcomes across cancer types.