Predicting Lung Health with High-Performance Machine Learning: Insights from Upper Respiratory Microbiome Biomarkers

BACKGROUND The major importance of upper respiratory tract (URT) microbiome diversity to the overall aerodigestive tract made it an ideal hotspot to determine lung health. However, the current sampling methods pose a bottleneck for large scale lung examination, as they are either invasive such as bronchoalveolar lavage (BAL), or inconsistent like sputum. As COVID-19 pushes innovation for mass surveillance, the self-collected gargle sampling method gained popularity as it is non-invasive, convenient, and requires minimum sampling skill. Combined by the power of sequencing technology and machine learning (ML) algorithms, gargle specimen analysis could be the novel approach for lung health surveillance. Here, we carried out shotgun metagenomics study to compare microbiome diversity between sputum and gargle specimens from 3 subject’s lung health groups, namely healthy, acute, and chronic. We also discovered biomarkers driving microbiome differences in healthy-chronic subjects and used that insight to develop ML based predictive models.RESULTS We found that influence of specimen types on microbiome diversity is significantly inconsequential compared to that of subjects’ health. Our biomarker analysis revealed higher commensals abundance in healthy samples, as well as pathogens predominance in chronic subjects. By using this insight, our prediction models achieve an excellent discriminatory power to distinguish healthy subjects from chronic patients.CONCLUSION Findings in our study demonstrated feasibility of ML-based gargle analysis as an alternative lung health surveillance approach.