Hierarchical Graph Learning with Multimodal Distillation for Breast Cancer Survival Prediction
In the authors' words
Imaging mass cytometry (IMC) and DNA methylation can provide complementary information for breast cancer prognosis, yet these measurements may be unavailable for new patients. In this paper, we propose HEAL-Surv, a teacher-student framework that uses both modalities during training to predict five-year net disease-specific survival (DSS) from clinical variables and molecular profiles at inference. An additive Cox teacher integrates all four modalities, using a hierarchical IMC graph encoder to capture local cell relationships and coarser tissue organisation. The student learns from observed outcomes and cross-fitted teacher survival curves, giving more weight to targets that agree across random initialisations. In five-fold cross-validation on 423 METABRIC patients, the full teacher achieved a mean five-year Uno's concordance of 0.723. Removing any modality reduced mean concordance and increased mean Brier-based errors. Relative to the matched student trained on observed outcomes alone, distillation increased mean five-year Uno's concordance from 0.658 to 0.704, with higher integrated and five-year Brier scores. In a separate external evaluation, an IMC-clinical variant trained on METABRIC achieved a five-year Uno's concordance of 0.660 on 265 Basel patients without refitting. These findings support multimodal supervision for improving risk discrimination with fewer inference inputs, while external validation of the student remains necessary.
Appeared: Wednesday, November 18. medRxiv. Preprint, not yet peer-reviewed.