Metabolomic machine learning predictor for the diagnosis of aspiration pneumonia

Abstract

Aspiration pneumonia (AP) is a common disease, particularly among the elderly, and is diagnostically challenging due to its non-specific presentation and lack of a diagnostic gold standard or reliable biomarkers. Metabolomics may enable accurate non-invasive diagnosis. To address this gap, this study aims to develop and validate metabolomic machine learning models for AP diagnosis within a multicenter prospective observational cohort. We enrolled 121 patients (Cohort 1) with AP or non-aspiration pneumonia (NonAP), nearly all of whom were from the intensive care unit. Non-targeted metabolomic profiling of serum, bronchoalveolar lavage fluid, and sputum was performed using liquid chromatography–mass spectrometry. Machine learning models were constructed for each biofluid. We also measured peripheral blood cytokines and built corresponding diagnostic models, against which the performance of the metabolomic models was compared. The optimal serum model was subsequently validated using targeted metabolomics in 175 patients. In Cohort 1, distinct metabolic profiles distinguished AP from NonAP patients across all sample types. The serum-based model demonstrated superior diagnostic performance (area under the receiver operating characteristic curve (AUROC) 0.973) compared to bronchoalveolar lavage fluid and sputum models. A cytokine-only model performed poorly (AUROC 0.700), and integrating cytokines with metabolomics did not improve accuracy. Consequently, the serum biomarker panel was advanced to targeted validation. In the 175 patients, a refined model based on four serum metabolites (Galactitol, PC(18:0/20:4(5Z,8Z,11Z,14Z)), Quinolinic acid, N6,N6,N6-Trimethyl-L-lysine) retained high diagnostic accuracy (AUROC 0.900) and effectively discriminated both acute and chronic AP subtypes (AUROCs 0.905 and 0.895, respectively). We developed a serum metabolomics-based model with high diagnostic accuracy for AP, showcasing strong potential for clinical diagnosis.