A Missing-Modality Robust Attention Fusion Network for Breast Cancer Survival Prediction Using Incomplete Clinical, Genomic and Histopathology Data

Abstract

Breast cancer is one of the most frequently diagnosed cancers worldwide, and accurate prognosis is essential for planning treatment intensity, follow-up schedules, and patient counselling. The survival outcome of a patient is influenced by many complementary factors, including clinical and pathological attributes, molecular alterations captured by gene expression, and the tissue morphology visible in histopathology images. Multimodal deep learning can combine these heterogeneous sources to produce more informative prognostic estimates than any single modality. However, in real hospital and research cohorts, many patients do not have every modality available because of cost, tissue availability, differences in clinical protocols, or incomplete records. Most existing multimodal models assume complete data and therefore either discard incomplete patients or fill the missing values with simple placeholders, which reduces the size of the usable cohort and may introduce bias. This paper proposes "A Missing-Modality Robust Attention Fusion Network for Breast Cancer Survival Prediction Using Incomplete Clinical, Genomic and Histopathology Data" to address this limitation. The framework uses modality-specific encoders, a learnable null representation for absent modalities, a masked cross-modal attention fusion mechanism, and modality dropout during training so that the model learns to make reliable predictions from any available subset of modalities. A Cox proportional hazards objective is used to estimate an individual risk score, which can then be used to stratify patients into risk groups and generate survival curves. The performance of the framework can be evaluated using the concordance index, time-dependent area under the curve, integrated Brier score, and Kaplan–Meier log-rank analysis under different missing-modality rates. The proposed approach aims to make multimodal prognosis practical for incomplete real-world data, although its predictions require independent clinical validation before use in medical decision-making.