[Breath metabolomic characteristics of schizophrenia and their potential for auxiliary diagnosis]

The clinical diagnosis of schizophrenia (SCZ) primarily relies on psychiatrists' comprehensive judgment based on symptom assessment, medical history collection, and mental status examination. However, in the early stages of the disease, in cases with atypical symptom presentations, or when differentiation from other psychiatric disorders is required, diagnosis may be influenced by patients' subjective reporting, symptom fluctuations, and differences in evaluators' clinical experience. Detection of exhaled volatile organic compounds (VOCs) is a noninvasive, rapid, and repeatable approach that can provide metabolic information related to physiological and pathological states. This study aims to compare the characteristics of exhaled VOCs between patients with SCZ and healthy controls (HCs) using a high-pressure photon ionization time-of-flight mass spectrometry (HPPI-TOFMS) platform combined with machine learning methods, and to evaluate their discriminative performance for the auxiliary diagnosis of SCZ.

An exploratory case-control study was conducted, including 34 patients with SCZ and 34 HCs. All patients with SCZ were diagnosed independently by 2 licensed psychiatrists according to the Diagnostic and Statistical Manual of Mental Disorders, fifth edition (DSM-5). During study design, participants in the 2 groups were matched by sex, age, and body mass index (BMI). Exhaled breath samples were collected from all participants within 30 minutes after waking in the morning. Participants consumed a light diet on the evening before sampling and avoided foods with strong odors or irritant properties. Fasting for at least 8 hours before sampling was required, and participants were instructed not to brush their teeth on the sampling day but only to rinse their mouths with purified water. Healthy controls stayed for at least 8 hours in the same ward environment as patients with SCZ on the night before sampling to minimize the impact of short-term environmental exposure differences. End-tidal breath was collected as the primary analytical sample using disposable polyether ether ketone (PEEK) sampling bags, and HPPI-TOFMS analysis was completed within 1-3 hours after collection. Raw mass spectrometry data were processed by denoising, baseline correction, peak position calibration, peak detection, peak alignment, background subtraction, and interval normalization. A VOC feature matrix with mass-to-charge ratios (m/z) ranging from 20 to 200 was subsequently constructed. Random forest (RF) and partial least squares-discriminant analysis (PLS-DA) were used to build binary classification models. Model performance was evaluated using receiver operating characteristic (ROC) curves, area under the curve (AUC), sensitivity, and specificity. Internal validation was performed using Monte Carlo cross-validation (MCCV), and model stability and non-randomness were further assessed through repeated training, prediction probability distributions, and permutation testing. Candidate VOC features associated with SCZ were identified through between-group differential analysis, volcano plots, hierarchical clustering heatmaps, and RF feature selection. Selected significant peaks were subjected to preliminary compound annotation.

The SCZ and HC groups had identical sex distributions (21 males and 13 females in each group). No significant differences were observed between the groups in age, BMI, or smoking status (all P>0.05). The RF model constructed from exhaled VOC features demonstrated discriminative ability between patients with SCZ and HCs, with an AUC of 0.900 (95% CI 0.812 to 0.973), a sensitivity of 82.4% (28/34), and a specificity of 82.4% (28/34) at the corresponding classification threshold. After repeated model construction using different numbers of features, the AUC remained stable between 0.857 and 0.900, and the ROC curves showed similar shapes across feature sets, suggesting relatively stable discriminative performance within the current internal validation framework. Based on 100 repeated cross-validation runs, the RF model achieved a mean accuracy of 0.816 with a concentrated accuracy distribution. In permutation testing, classification performance of the randomly labeled models was significantly inferior to that of the original model (empirical P<0.001). PLS-DA model yielded an AUC of 0.816 (95% CI 0.696 to 0.953), with a sensitivity of 82.4% and a specificity of 58.8%, indicating inferior overall performance compared with the RF model. Through differential analysis, volcano plots, hierarchical clustering heatmaps, and RF feature selection, 10 candidate VOC features associated with SCZ were identified, corresponding to m/z values of 97.1, 119.1, 78.1, 115.1, 108.1, 73.1, 55.1, 56.1, 75, and 123.1. Among these, m/z 78.1, 115.1, and 108.1 were preliminarily annotated as benzene, 2-heptanone, and p-cresol, respectively. Reliable preliminary annotations for m/z 97.1 and 119.1 could not be established based on the available analytical information.

Exhaled VOC analysis combined with machine learning can identify discriminative breath metabolomic features between patients with SCZ and HCs. Under the current sample size and internal validation framework, the RF model demonstrated promising discriminative performance and outperformed the linear PLS-DA model. These findings suggest that SCZ-related breath metabolomic information may be characterized more by multi-feature patterns than by isolated alterations in individual VOCs. The candidate VOC features and qualitative results in this study may provide valuable clues for future breath metabolomics research in schizophrenia.