Integrating multimodal data through interpretable heterogeneous ensembles.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Data Science

Background:

• Multimodal data integration is crucial for predicting biomedical outcomes.
• Existing methods struggle with heterogeneous semantics and may lose local information.
• Late integration approaches are understudied in biomedical applications.

Purpose of the Study:

• To introduce Ensemble Integration (EI), a systematic late integration approach for multimodal biomedical data.
• To evaluate EI's performance in predicting protein function and COVID-19 mortality.
• To develop a novel interpretation method for EI models.

Main Methods:

• EI infers local predictive models from individual data modalities.
• Heterogeneous ensemble algorithms integrate local models into a global predictive model.
• EI was tested on STRING protein data and electronic health records for COVID-19 mortality.

Main Results:

• EI significantly improved prediction accuracy compared to individual modalities.
• EI outperformed established early integration methods for both tested problems.
• EI model interpretation identified key disease-relevant features for COVID-19 mortality.

Conclusions:

• The EI framework is effective for multimodal biomedical data integration and predictive modeling.
• EI offers a robust approach to leveraging diverse data sources for enhanced biomedical insights.
• The developed interpretation method aids in understanding EI model predictions.