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

Area of Science:

• * Medical Diagnostics
• * Computational Biology
• * Infectious Diseases

Background:

• * Prompt diagnosis of pulmonary infections in critically ill patients is challenging due to limitations in current diagnostic methods.
• * Metatranscriptomic sequencing is powerful but often lacks the timeliness required for clinical settings.
• * Rapid and cost-effective diagnostic tools are crucial for managing severe infections.

Purpose of the Study:

• * To develop a rapid and inexpensive adjunctive diagnostic strategy for pulmonary infections in severe patients.
• * To identify host gene expression signatures associated with different infection types.
• * To create accurate machine learning-based diagnostic models using limited gene expression data.

Main Methods:

• * Metatranscriptomic sequencing of bronchoalveolar lavage fluid (BALF) from critically ill patients.
• * Screening of characteristic host genes by comparing infected versus non-infected patient data.
• * Construction of ensemble machine learning models (including Lasso regression) using identified gene signatures.

Main Results:

• * Machine learning models achieved high accuracy in distinguishing infection from non-infection (AUC=0.984), bacterial infection (AUC=0.98), and viral infection (AUC=0.98) during cross-validation.
• * Test cohorts showed consistent accuracy, discerning infection status (AUC=0.865) and type (viral: AUC=0.934, bacterial: AUC=0.871).
• * The developed diagnostic strategy demonstrated accuracy comparable to metatranscriptomic sequencing.

Conclusions:

• * A rapid, cost-effective diagnostic strategy using host gene expression signatures was developed.
• * The method enables timely identification of infection status and type in pulmonary infections.
• * This approach serves as a valuable adjunctive tool for clinical decision-making in severe cases.