Stacking ensemble learning models diagnose pulmonary infections using host transcriptome data from metatranscriptomics

The prompt diagnosis of pulmonary infections with unknown etiology in patients in severe condition remains a challenge due to the lack of rapid and effective diagnostic methods. While metatranscriptomic sequencing offers a powerful approach, its clinical utility is often limited by issues of timeliness. In this study, we conducted metatranscriptomic sequencing on bronchoalveolar lavage fluid (BALF) collected from critically ill, severely ill, and ICU patients. Based on microbial detection results, patients were classified into four types: negative, bacterial infection, viral infection, and fungal infection. To identify host gene expression signatures associated with infection, we screened characteristic genes from human metatranscriptomic data by comparing 70% of patients with confirmed infections vs. non-infections. Leveraging these characteristic genes, we constructed classification sub-models employing 13 types of machine learning algorithms, and we further integrated these sub-models into stacking-based ensemble models with Lasso regression, resulting in diagnostic models that required only a small set of gene expression inputs. The average performance of five-fold cross-validation demonstrated high diagnostic accuracy: distinguishing infection from non-infection (AUC = 0.984), bacterial infection from non-bacterial infection (AUC = 0.98), and viral infection from non- viral infection (AUC = 0.98). Test cohorts' results demonstrated the method's high diagnostic accuracy consistency with metatranscriptomic sequencing in discerning patient infection status (AUC = 0.865) and the type of infection (viral: AUC = 0.934, bacterial: AUC = 0.871). Our study presented a rapid and inexpensive adjunctive diagnostic strategy that achieves diagnostic accuracy comparable to metatranscriptomic sequencing, enabling timely identification of both infection status and type in pulmonary infections.