Hierarchical Graph Learning with Multimodal Distillation for Breast Cancer Survival Prediction

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.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-17
DOI
https://doi.org/10.64898/2026.11.14.26368121
Primary Topic
Breast Cancer Treatment Studies
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Hierarchical Graph Learning with Multimodal Distillation for Breast Cancer Survival Prediction

Văn Ân Nguyễn, Hai Nam Tran, Dang Hung Tran, Minh Hai Nguyen Ngoc et al.
4 links
medRxiv
Breast Cancer Treatment Studies
preprint

Hierarchical Graph Learning with Multimodal Distillation for Breast Cancer Survival Prediction

Văn Ân Nguyễn, Hai Nam Tran, Dang Hung Tran, Minh Hai Nguyen Ngoc, Dang Quynh Anh, Binh Minh Tran
preprint en

Abstract

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.

medRxiv
The University of Sydney (AU), Hanoi Medical University (VN), UNSW Sydney (AU), Hanoi Medical University Hospital (VN), Hanoi University of Industry (VN)
Good health and well-being
Breast Cancer Treatment Studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon