Exhaled breath performance in diagnosing ischemic heart disease utilizing a machine learning model

Ischemic heart disease (IHD) remains the leading global cause of mortality and morbidity. Current diagnostic approaches face significant limitations in accessibility, cost, invasiveness, and accuracy, creating an urgent need for innovative non-invasive screening methods. To evaluate the ability of a machine learning (ML) model based on dynamic exhaled breath volatile organic compound (VOC) patterns to detect stress-induced myocardial perfusion defects, used here as the study reference outcome related to IHD. This prospective single-center study enrolled 80 participants (31 with stress-induced myocardial perfusion defects confirmed by multidetector computed tomography with perfusion assessment and 49 controls). All participants underwent real-time breath analysis using PTR-TOF-MS-1000 at rest and after bicycle ergometry stress testing. ML models were developed using delta changes in VOC patterns between baseline and post-exertion measurements, with rigorous leave-one-out cross-validation and comprehensive performance metrics. The weighted ensemble classifier achieved an AUC of 0.743 (95% CI: 0.622-0.840), with a sensitivity of 0.774 (95% CI: 0.600-0.905) and a specificity of 0.633 (95% CI: 0.500-0.760). Feature importance analysis identified specific VOCs, particularly those with m/z 94.053731, 90.951122, and 60.055305, which demonstrated consistent diagnostic significance across different temporal measurement points (delta10). This study demonstrates the feasibility of dynamic exhaled breath analysis combined with ML for the detection of IHD. However, the present findings should be considered preliminary and exploratory, as the model was validated internally only. Given the current sensitivity, the approach is not suitable as a standalone screening tool, but may serve as a non-invasive adjunctive or triage-support tool. External validation in independent cohorts is required before clinical implementation.