Uncertainty-Aware Survival Analysis With Dirichlet Distribution for Multi-Scale Pathology and Genomics.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Biostatistics

Background:

• AI in digital pathology has advanced survival prediction.
• Current survival analysis methods often discretize time, ignoring uncertainty and patient heterogeneity.
• Censored data in survival analysis amplifies uncertainty and variability.

Purpose of the Study:

• To develop a novel survival analysis framework addressing limitations of existing methods.
• To enhance uncertainty awareness in survival prediction models.
• To integrate multi-modal data, including pathology images and genomic data, for improved survival analysis.

Main Methods:

• Utilized the Dirichlet distribution to model discretized outputs as continuous probability distributions, enhancing uncertainty representation.
• Developed a universal multi-modal survival analysis loss function based on uncertainty-driven fusion.
• Proposed the Uncertainty-Aware Multi-Modal Survival Analysis (UMSA) framework to analyze interactions between multi-scale pathological images and genomic data.

Main Results:

• The UMSA framework demonstrated state-of-the-art performance in survival prediction tasks.
• Experimental evaluations on five public datasets validated the effectiveness and scalability of the proposed approach.
• The method provides more accurate representations of uncertainty in survival prediction.

Conclusions:

• The UMSA framework offers a significant advancement in multi-modal survival analysis by incorporating uncertainty awareness.
• This approach effectively leverages pathological images and genomic data for more robust survival predictions.
• UMSA shows promise for improving clinical decision-making through enhanced survival prediction accuracy.