Machine learning-based classification of COVID-19 severity using respiratory microbiome profiles from shotgun...

Accurate clinical triage is critical for optimizing decision-making and resource allocation during infectious disease outbreaks such as COVID-19. In this study, we present an AI-driven decision-support tool for the triage of COVID-19 patients based on respiratory microbiome profiles derived from shotgun metagenomic sequencing. We analyzed 477 shotgun respiratory metagenomes from three independent public cohorts and generated genus-level taxonomic profiles, which were integrated with minimal clinical metadata (age, sex, and antibiotic exposure) to train supervised machine-learning models, including Random Forest, Support Vector Machine, and XGBoost. Model performance was evaluated using standard classification metrics, cross-validation, and particle swarm optimization for hyperparameter tuning. Across cohorts, we observed a consistent transition from microbiomes dominated by commensal taxa to dysbiotic states enriched in opportunistic and clinically relevant genera, particularly Acinetobacter and Staphylococcus, in severe and deceased patients. Among the evaluated models, XGBoost consistently achieved the best performance, reaching up to 96.1% accuracy, 97.6% F1-score, and 98.2% ROC-AUC in individual cohorts. When trained on the integrated dataset, XGBoost maintained robust performance (95.1% accuracy, 97.2% F1-score, 94.3% ROC-AUC) and demonstrated greater stability and lower variance compared to alternative models. Feature-importance analyses identified a compact and interpretable set of recurrent microbial predictors, and reduced-feature models retained substantial discriminative power when augmented with key clinical variables. These results support the respiratory microbiome as a valuable source of information for outcome-oriented clinical triage and position microbiome-informed machine learning as a scalable and interpretable decision-support approach for managing COVID-19 and future infectious disease scenarios.