Lung Microbiome Signatures and Explainable Predictive Modeling of Glucocorticoid Response in Severe Community Acquired Pneumonia

Yeong-Nan Cheng^1,2,3Guan-Ting Chen^1,2,3Wei Chih Huang^1,2,3Yen-Peng Chiu^1,4Yun Tang^1,2,3Pin-Kuei Fu^5,6*Tzong-yi Lee^1,2,3*

• ¹National Yang Ming Chiao Tung University, Hsinchu, Taiwan
• ²National Yang Ming Chiao Tung University College of Engineering Bioscience, Hsinchu, Taiwan
• ³National Yang Ming Chiao Tung University Institute of Bioinformatics and Systems Biology, Hsinchu, Taiwan
• ⁴Institute of Data Science and Engineering, College of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
• ⁵Taichung Veterans General Hospital, Taichung, Taiwan
• ⁶National Chung Hsing University Department of Post-Baccalaureate Medicine, Taichung, Taiwan

Systemic glucocorticoids (SG) are administered to quell hyper-inflammation in severe community acquired pneumonia (SCAP), yet trials report inconsistent efficacy and no mechanistic explanation. We enrolled 200 ventilated SCAP patients, whom received hydrocortisone within 48 h of ICU admission, and generated longitudinal lower-airway microbiome profiles by 16S rRNA amplicon and metagenomic sequencing on ICU Days 1, 3 and 7. Compositional data were integrated with clinical variables through a fully reproducible bioinformatics analysis workflow. Baseline community structures did not differ between SG and control cohorts, but by Day 7 survivors exhibited enrichment of Actinobacteria and Gammaproteobacteria whereas non-survivors accumulated Alphaproteobacteria and Campylobacteria. A random-forest model restricted to Bacilli and Alphaproteobacteria achieved AUROC = 0.89 (sensitivity 0.83, specificity 0.81) on a patient-held-out test set, significantly outperforming conventional severity indices like APACHE II, SOFA and mNUTRIC scores. Collectively, our results demonstrate that SG therapy imposes reproducible ecological pressures on the lung microbiome and that a two-feature microbial fingerprint can forecast treatment success with single-sample resolution. These findings show that SG therapy actively reshapes the respiratory ecosystem and that lightweight microbiome-aware machine learning can stratify treatment response, offering a tractable path toward precision corticosteroid stewardship.