Key Findings
- Molecular features were the strongest contributors to prognostic accuracy for early distant recurrence (within five years), while histopathology features added meaningful prognostic value for late distant recurrence.
- Multimodal AI models that integrate image, clinical, and molecular inputs captured these complementary signals, with the ICM+ model showing the best overall accuracy for individualized recurrence risk assessment.
- The prognostic accuracy of the optimal ICM+ and CM+ models was independently validated in a holdout validation set from the TAILORx trial.