Multimodal data fusion using sparse canonical correlation analysis and cooperative learning: a COVID-19 cohort study

Area of Science:

• Computational biology
• Medical informatics
• Data science

Background:

• High-dimensional biomedical data offers insights into disease phenotypes and outcomes.
• Analyzing multimodal data (genomics, imaging, clinical, lab) presents significant challenges.
• COVID-19 patient data requires advanced analytical techniques for comprehensive understanding.

Purpose of the Study:

• To analyze multimodal data from a COVID-19 patient cohort using unsupervised and supervised sparse linear methods.
• To identify relationships between different data modalities and predict clinical outcomes.
• To explore viral genome encoding for variant classification and phylogenetic analysis.

Main Methods:

• Prospective cohort study of 149 adult COVID-19 patients.
• Sparse Canonical Correlation Analysis (CCA) for cross-modal data relationships.
• Cooperative learning for predicting Intensive Care Unit (ICU) admission.
• Word2Vec natural language processing for viral genome encoding.

Main Results:

• Serum biomarkers correlated with radiomics features (cor=0.596, p<0.001).
• Unsupervised analysis revealed distinct clinical phenotypes.
• Word2Vec encoding separated SARS-CoV-2 variants and preserved phylogenetic relationships.
• A quadruple model achieved an Area Under the Curve (AUC) of 0.87 and accuracy of 0.77 for outcome prediction.

Conclusions:

• Sparse CCA and cooperative learning are powerful for high-dimensional, multimodal data analysis.
• These methods facilitate investigation of multivariate associations in unsupervised and supervised tasks.
• The approach provides a framework for understanding complex diseases like COVID-19.