Robust Multimodal Fusion for Survival Prediction in Cancer Patients.

Area of Science:

• Oncology
• Bioinformatics
• Artificial Intelligence

Background:

• Multimodal deep learning models offer potential for enhanced cancer patient survival predictions and treatment planning.
• Existing models often show limited improvement over unimodal approaches, necessitating robust validation of multimodal efficacy.

Purpose of the Study:

• To introduce the Robust Multimodal Survival Model (RMSurv), a novel discrete late fusion model for improved cancer survival prediction.
• To demonstrate substantial and consistent advantages of multimodal data integration over unimodal models.

Main Methods:

• RMSurv employs a discrete late fusion technique with synthetic data generation for time-dependent modality weighting.
• The model integrates up to six data modalities from TCGA non-small cell lung cancer and pan-cancer datasets.
• A novel statistical feature normalization enhances interpretability and accuracy of discrete survival predictions.

Main Results:

• RMSurv achieved a 0.0273 higher Concordance Index (C-Index) than the best unimodal model on the TCGA LUAD dataset.
• Outperformed existing early and late fusion methods by significant margins.
• Demonstrated superior performance on combined TCGA non-small-cell lung cancer and pan-cancer datasets.

Conclusions:

• RMSurv establishes a new benchmark for survival prediction models, showcasing robust multimodal benefits.
• The model's advancements highlight its potential for powerful survival prediction in pan-cancer settings.
• Consistent and substantial improvements validate the efficacy of multimodal data in cancer research.