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

The human microbiome, a complex ecosystem of microorganisms residing in different body compartments, significantly influences health outcomes and disease progression, however, leveraging this data for developing clinical prediction models remains challenging due to its high dimensionality, sparsity, and compositional nature. Traditional machine learning approaches often struggle to capture the intricate microbial interactions that contribute to mortality risk, particularly when analyzing data across multiple compartments with distinct microbial compositions. To address these limitations, we introduce a novel framework utilizing an autoencoder-based model trained on high-dimensional microbiome data collected from oral, lung, and gut compartments. Our approach encodes microbiome data into a low-dimensional latent space while preserving essential microbial community characteristics, enabling more effective feature extraction and pattern recognition than conventional dimensionality reduction techniques. Through systematic evaluation of three data configurations—microbiome taxa only, clinical data only, and an integrated model combining both—we demonstrated that the integrated approach consistently achieved superior prediction accuracy (98 % in lung microbiome) compared to using either data source independently. Clinical data alone provided reasonable but inconsistent performance (70–90 %), while microbiome taxa alone yielded the weakest results (53–65 %). Furthermore, our investigation of preprocessing techniques revealed that applying z-score normalization to the taxa data significantly enhanced performance and substantially improved recall metrics across all compartments. By analyzing compartment-specific microbial contributions, our study reveals distinct predictive roles of the oral and lung microbiomes compared to the gut microbiome, underscoring of body-site specificity in microbiome-based predictive modeling.