Machine learning-based prediction model for bronchoalveolar lavage efficacy in severe pneumonia

Area of Science:

• Pulmonary Medicine
• Critical Care Medicine
• Medical Informatics

Background:

• Community-acquired severe pneumonia poses a significant clinical challenge.
• Bronchoalveolar lavage (BAL) is a treatment modality, but predicting its efficacy is crucial for patient management.
• Identifying factors influencing BAL treatment outcomes can optimize patient care.

Purpose of the Study:

• To develop and validate a predictive model for the clinical efficacy of BAL in severe pneumonia.
• To identify key indicators of inflammatory response and blood gas analysis that predict BAL treatment outcomes.
• To utilize machine learning algorithms for robust predictive modeling.

Main Methods:

• A cohort of 206 severe pneumonia patients undergoing BAL was divided into training (n=144) and validation (n=62) sets.
• Univariate analysis, LASSO, and multivariate logistic regression identified independent risk factors for poor efficacy.
• Random Forest (RF), K-nearest neighbor (KNN), and Gradient Boosting (GB) models were constructed and evaluated using ROC AUC.

Main Results:

• Independent risk factors for poor BAL efficacy included advanced age (≥60 years), COPD, elevated procalcitonin (PCT ≥2 ng/mL), high C-reactive protein (CRP ≥100 mg/L), and impaired oxygenation (PaO₂ <60 mmHg) or elevated CO₂ (PaCO₂ ≥50 mmHg).
• The RF model demonstrated superior performance with AUCs of 0.799 (training) and 0.778 (validation), outperforming KNN and GB models.
• Poor clinical efficacy was observed in 22.22% of the training set and 20.97% of the validation set.

Conclusions:

• A validated Random Forest-based predictive model effectively identifies factors influencing BAL treatment outcomes in severe pneumonia.
• The model provides a foundation for hypothesis testing and clinical prediction research in severe pneumonia.
• Further multicenter external validation and inclusion of a non-BAL control group are necessary for clinical application in individualized treatment planning.