Integrating multi-compartment microbiome data with clinical parameters enhances mortality prediction using...

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• The human microbiome is crucial for health but challenging to model due to high dimensionality and complexity.
• Traditional machine learning methods struggle with intricate microbial interactions for mortality risk prediction.

Purpose of the Study:

• To develop a novel framework for predicting mortality risk using multi-compartment human microbiome data.
• To enhance feature extraction and pattern recognition in high-dimensional microbiome data.

Main Methods:

• Utilized an autoencoder-based model for encoding microbiome data into a low-dimensional latent space.
• Evaluated three data configurations: microbiome taxa only, clinical data only, and integrated data.
• Investigated the impact of z-score normalization preprocessing on taxa data.

Main Results:

• The integrated model combining microbiome and clinical data achieved superior prediction accuracy (98% in lung microbiome).
• Clinical data alone showed inconsistent performance (70-90%), while microbiome data alone was weakest (53-65%).
• Z-score normalization significantly improved performance and recall metrics across all compartments.

Conclusions:

• An integrated approach using multi-compartment microbiome and clinical data offers superior mortality risk prediction.
• Body-site specificity of the microbiome plays a distinct role in predictive modeling.
• Autoencoder models provide effective dimensionality reduction for complex microbiome datasets.