Multimodal Genotype and Phenotype Data Integration to Improve Partial Data-Based Longitudinal Prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Medical imaging analysis

Background:

• Multimodal data analysis is crucial in bioinformatics but limited by the need for all data during prediction.
• Real-world biomedical applications often face missing data due to high collection costs.
• Predicting longitudinal disease progression requires assessing current status and future outcomes for early intervention.

Purpose of the Study:

• To develop a novel adversarial mutual learning framework for longitudinal disease progression prediction.
• To enable performant models that utilize a single data modality for prediction, addressing the missing modality problem.
• To simultaneously assess current disease severity and forecast future progression for chronic diseases.

Main Methods:

• A single-modal model learns from a pretrained multimodal model via adversarial mutual learning.
• The framework infers auxiliary modality representations from the main modality's representations.
• Combines representations to predict longitudinal outcomes, trained using the Age-Related Eye Disease Study dataset.

Main Results:

• The proposed method outperforms baseline approaches in classifying current disease severity.
• The framework demonstrates superior effectiveness in forecasting future disease severity.
• Successful application to retinal imaging genetics for age-related macular degeneration (AMD) early diagnosis.

Conclusions:

• The adversarial mutual learning framework effectively handles missing modalities in longitudinal disease prediction.
• The method enables accurate simultaneous assessment of current and future disease status.
• This approach holds promise for early diagnosis and intervention strategies in chronic diseases.