A cancer-type-aware framework for robust multimodal survival prediction under missing modalities

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Multimodal cancer prognosis faces challenges with incomplete data, lack of cancer-specific biological modeling, and cross-institutional instability.
• Existing methods struggle to integrate diverse data types (histopathology, RNA expression, clinical text) effectively in real-world scenarios.

Purpose of the Study:

• To develop and validate a cancer-type-aware framework for robust multimodal survival prediction.
• To address data incompleteness, cancer heterogeneity, and cross-institutional variability in cancer prognosis.

Main Methods:

• A novel framework employing adaptive gated fusion for missing data imputation and a hybrid architecture for cancer heterogeneity.
• Histopathology utilized as an anchor modality, adaptively integrating RNA expression and clinical text.
• Evaluation across 10 The Cancer Genome Atlas (TCGA) cancer types and cross-institutional validation.

Main Results:

• Superior performance in 10 TCGA cancer types (C-indices 0.578-0.778), achieving state-of-the-art in six.
• Maintained predictive performance with missing RNA (C-indices 0.621-0.627) and clinical text (C-indices 0.568-0.606).
• Demonstrated robust cross-institutional performance (SD <0.040 in 8/10 types).

Conclusions:

• The proposed framework successfully handles missing data, cancer heterogeneity, and cross-institutional stability for multimodal survival prediction.
• Provides a robust computational foundation for integrating diverse data types in cancer prognosis.
• Paves the way for future prospective clinical validation of multimodal prognostic models.