Predicting and diagnosing pneumonia in patients undergoing elective cardiac surgery via machine learning analysis of exhaled volatile carbonyl compounds.

Objective: Pneumonia remains one of the most common postoperative complications after elective cardiac surgery. Early intervention could lead to improved patient outcomes, including lower rates of intensive care unit admissions, and shorter hospital stays. Volatile organic compounds (VOCs) in exhaled breath have shown promise in diagnosis and classification for various lung-related conditions. The study aims to diagnose and predict the onset of pneumonia in patients undergoing elective cardiac surgery via machine learning (ML) analysis of VOCs.

Methods: Patients undergoing elective cardiac surgery (n = 75) were enrolled in the study (March 2023 to July 2024). Each patient's breath was collected in a 600-mL Tedlar bag preoperatively, within 24 hours, and every 3 days. The pneumonia group consisted of those who developed clinical signs of pneumonia postoperatively. Carbonyl compounds in the breath were captured on a microchip and identified using mass spectrometry. An ML workflow was implemented to build a model for pneumonia diagnosis (trained on pre- and postoperative VOC samples) and to build a prediction model of pneumonia development (trained on preoperative samples) (alpha 0.05).

Results: Of the 75 patients enrolled during the study period, 10 developed clinical signs of pneumonia. The majority of patients had undergone coronary artery bypass grafting (50.1%), followed by aortic valve/root replacement (22.7%), concomitant coronary artery bypass grafting and valve (16%), and mitral valve repair/replacement (8%). Twenty-four carbonyls were selected by the pneumonia diagnosis model, including formaldehyde, hexanal, CHO, CHO, hexanone, and hydroxy-butanal. The proposed pneumonia diagnosis model had an area under the receiver operating characteristic of 0.833 and an area under the precision-recall curve of 0.818 on the test set. In contrast, 4 carbonyls (heptanal, octenone, CHO, and acetone) were selected by the model to predict the onset of pneumonia using preoperative breath samples (area under the receiver operating characteristic of 0.833 and area under the precision-recall curve of 0.818 on the test set).

Conclusions: This pilot study demonstrates that VOCs captured from breath can be used to train and test ML models for diagnosis and prediction of pneumonia onset in patients undergoing elective cardiac surgery. This finding has implications for guiding perioperative and postoperative strategies for preventing pneumonia.