Abstract PD11-11: Multimodal Deep Learning for Recurrence Stratification for Early-Stage Breast Cancer in Resource-Constrained Environments

Key Points

• The aim is to validate TRINITY AI for predicting recurrence risk in early-stage breast cancer using multimodal data.
• TRINITY AI was trained on 1219 breast cancer cases from TCGA and CPTAC.
• A self-supervised transformer model inferred transcriptomic profiles from H&E whole slide images.
• Clinical variables including patient age, tumor size, and grade were integrated into a shared latent space.
• The model was validated using an external cohort of 166 cases with matched Oncotype Dx scores.
• Prognostic utility was evaluated in 1051 TCGA cases using multivariable Cox models.
• TRINITY AI classified low-risk disease with 95% specificity and 92% negative predictive value.
• Outperformed traditional clinicopathologic nomograms with an area-under-the-curve of 0.88.
• Achieved a negative predictive value greater than 87% and specificity above 92% in cohort-specific analysis.
• High-risk predictions associated with a 3.80-fold increase in distant recurrence hazard after adjustments.
• C-index of 0.698 indicates its utility in predicting distant breast cancer recurrence.