Deep survival analysis in multimodal medical data: a parametric and probabilistic approach with competing risks

Area of Science:

• Oncology
• Bioinformatics
• Artificial Intelligence

Background:

• Accurate cancer survival prediction is crucial for patient prognosis and treatment planning.
• Traditional methods using single data types struggle to capture tumor complexity.
• Integrating multiple data sources offers a more comprehensive approach to survival analysis.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning framework for enhanced cancer survival prediction.
• To assess the impact of integrating diverse data modalities (clinical, molecular, imaging) on survival analysis.
• To address both single and competing risks scenarios in cancer prognosis.

Main Methods:

• Proposed SAMVAE (Survival Analysis Multimodal Variational Autoencoder), a deep learning architecture integrating six data modalities.
• Utilized modality-specific encoders to project data into a shared latent space for robust prediction.
• Evaluated the model on breast cancer and lower-grade glioma cohorts with tailored data preprocessing and optimization.

Main Results:

• Demonstrated successful integration of multimodal data for standard and competing risks survival analysis.
• Achieved competitive predictive performance against state-of-the-art multimodal survival models.
• Presented a unified probabilistic multimodal fusion framework for continuous-time competing risks.

Conclusions:

• SAMVAE offers a powerful tool for personalized survival curve generation from heterogeneous data.
• The framework provides clinically meaningful statistics and patient-specific insights.
• This probabilistic multimodal fusion approach advances risk modeling and interpretable survival analysis in oncology.