Skip to main content
Back to timeline
medRxivSource publication:

Breast cancer five-year survival can be predicted without tissue imaging: HEAL-Surv lifts concordance from 0.658 to 0.704 by distilling a teacher

Synopsis

HEAL-Surv puts imaging mass cytometry (IMC, which measures protein markers per cell while keeping cell positions) and DNA methylation into a teacher used only during training, so a student predicts five-year net disease-specific survival in breast cancer from molecular and clinical inputs alone at inference. In five-fold cross-validation on 423 METABRIC patients, distillation raised the matched student's mean five-year Uno's concordance from 0.658 to 0.704, while the integrated and five-year Brier scores both rose.

Source-provided article image: Hierarchical Graph Learning with Multimodal Distillation for Breast Cancer Survival Prediction
Fig. 1 · medRxiv · Page 4

Interpretation

When tissue-level IMC and methylation measurements are unavailable for a new patient, HEAL-Surv keeps those modalities in training only and leaves molecular profiles plus clinical variables at inference. Earlier survival models represented each patient as one feature vector, so once tissue measurements were aggregated into a single vector, which cells were neighbours and how those neighbourhoods were organised was lost. In five-fold cross-validation on 423 METABRIC patients, the four-modality teacher reached a mean five-year Uno's concordance of 0.723, and removing any modality lowered mean concordance and raised mean Brier-based errors.

The student learns more about risk ranking from the teacher's survival curves than from observed outcomes alone, and needs neither the teacher, IMC nor methylation at inference. The matched student with the same architecture and inference inputs, trained on observed outcomes alone, had a mean five-year Uno's concordance of 0.658. Adding a curve-matching loss weighted by agreement across random initialisations raised that student's mean to 0.704, close to DMGN's 0.701, while the integrated Brier score rose from 0.071 to 0.074 and the five-year Brier score from 0.142 to 0.147.

The teacher's IMC branch represents tissue as a cell graph, learning local cell adjacency first and then coarsening it into broader tissue regions. The non-hierarchical teacher keeps the same cell graphs and cell-level message passing but drops the two-stage pooling and reads out each tissue core directly. On the same 423 patients and five folds, the hierarchical teacher's mean five-year Uno's concordance rose from 0.618 for the non-hierarchical version to 0.723, with smaller fold SDs on both concordance indices.

Perspective

The demonstrated scope is five-fold cross-validation on 423 METABRIC patients plus an external evaluation on 265 Basel patients without refitting; the external evaluation covers the IMC-clinical variant, with a five-year Uno's concordance of 0.660, while the full teacher and the molecular-clinical student still require external validation.

Distillation improved risk-ranking concordance, but the integrated and five-year Brier scores rose at the same time, so probability error did not improve in step; the authors note that distillation from teachers with different input sets would help quantify how much IMC and methylation each contribute to the ranking gain, and external validation and direct calibration assessment of the student remain priorities.