Abstract :

Breast Cancer Prognosis Requires The Combined Interpretation Of Clinical Attributes, Molecular Profiles, And Pathological Findings, Because No Single Data Source Fully Describes The Biological And Clinical Behaviour Of A Tumour. Multimodal Deep Learning Has Shown Promise In Integrating These Sources, But Two Practical Issues Limit Its Adoption. First, Real-world Patient Records Are Often Incomplete, With One Or More Modalities Missing For A Considerable Proportion Of Patients. Second, Deep Multimodal Models Are Frequently Difficult To Interpret, Which Reduces Clinical Trust In Their Predictions. This Paper Proposes "An Explainable Patient-Similarity Graph Neural Network For Breast Cancer Recurrence And Survival Prognosis With Missing Multimodal Data". In The Proposed Framework, Each Patient Is Represented As A Node In A Population Graph, And Edges Connect Patients Who Are Similar On The Modalities That Both Of Them Have Available. Modality-specific Encoders Generate Masked Embeddings That Exclude Missing Data, And A Graph Attention Network Propagates Information Between Similar Patients, Allowing A Patient With A Missing Modality To Benefit From The Evidence Of Comparable Patients Rather Than From Artificial Placeholder Values. A Multi-task Prediction Module Jointly Estimates Recurrence Status And A Survival Risk Score, And An Explanation Module Combines SHAP Feature Attributions, Modality Contribution Analysis, And Neighbour Attention Weights To Show Which Features And Which Similar Patients Influenced Each Prediction. The Framework Can Be Evaluated Using The Concordance Index, Area Under The Receiver Operating Characteristic Curve, F1-score, Kaplan– Meier Risk Stratification, And Explanation Consistency Under Varying Missing-data Rates. The Proposed Approach Aims To Provide Robust And Transparent Prognostic Support, Although Its Predictions Require Independent Clinical Validation Before Use In Medical Decision-making.

Published:

19-10-2026

Issue:

Vol. 26 No. 10 (2026)

Page Nos:

81-88

Section:

Articles

License:

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

How to Cite