This cross-sectional study, conducted from November 2021 to January 2022 at the East Division of the First Affiliated Hospital of Sun Yat-sen University in Guangzhou, China, received approval from the Ethics Committee of the First Affiliated Hospital of Sun Yat-sen University (No. 2022-016). All subjects provided signed informed consent. Inclusion criteria were adult participants aged 18–80 years with clinical suspicion of malignant thoracic tumors, supported by imaging evidence (CT/PET-CT) and a multidisciplinary team (MDT) assessment prior to histological confirmation. Eligible participants encompassed treatment-naïve, newly diagnosed thoracic cancer patients scheduled for surgical resection for diagnostic evaluation, and patients with a history of treated or recurrent thoracic malignancy, provided comprehensive treatment records were available. Exclusion criteria encompassed individuals who were unwilling or unable to provide in-person informed consent, those with unqualified breath samples, patients with relapsed diseases and incomplete treatment histories, individuals suffering from other malignant tumors, those with severe bronchial asthma or confirmed tuberculosis, and those with severe liver damage or kidney diseases. Each participant had undergone resection surgery and was pathologically confirmed to be categorized into one of the following groups: lung cancer, thymoma, esophageal cancer, and benign disease controls. Demographic and clinical information were meticulously recorded and collected. This study was registered in the Chinese Clinical Trial Registry (Registration No.: ChiCTR2200061264).

All samples were collected following the same standardized procedure. Prior to collection, subjects were asked to rinse their mouths with purified water and rest for 15 minutes to stabilize their respiratory patterns. All subjects were required to abstain from food and beverages except water and smoking for at least 12 hours before the collection. To minimize the influence of diurnal metabolic variations, all collections were scheduled between 7:00 AM and 9:00 AM.

Subjects were instructed to remain seated and breathe normally through a mask for 3 minutes. During exhalation, breath samples were concurrently drawn through a breath sampler (CXBC-Alpha, ChromX Health Co., Ltd) containing an internal sampling pump and a flow control module ( Figure 1 ). 900 mL of breath samples were collected at a rate of 300 mL/min and directed into thermal desorption tubes. These tubes, pre-conditioned with 99.9% nitrogen gas to ensure a clean and inert environment, contained Carbopack X and Carbopack B for sample enrichment, concentrating the target compounds for later analysis. All collected samples were sealed with inert end caps immediately and stored at -20°C to maintain their integrity and analyzed by thermal desorption-gas chromatography-mass spectrometry (TD-GC-MS) within 7 days to ensure timely and accurate results.

Breath samples were analyzed by TD-GC-MS using a system incorporating a high-throughput autosampler, a thermal desorber (TD100-xr, MARKES), and an 7890B-5977A GC/MSD (Agilent Technologies). Separation was performed on an HP-5MS capillary column with nitrogen carrier gas. The mass spectrometer operated in electron ionization (EI) mode at 70 eV, acquiring data in full scan mode (m/z 33-450). Detailed instrument parameters are provided in Supplementary Materials.

Raw GC-MS data were processed using MSDial v5.4 for peak detection, quantification, and alignment. The software generated matrices of peak area (VOC area matrix) and signal-to-noise ratio (SNR matrix). Prior to statistical analysis, a data preprocessing and filtering protocol was implemented in Python 3.9.18 to ensure data robustness. Firstly, the signal-to-noise ratio (SNR) matrix was used to assess response reliability. A VOC measurement was classified as valid if its SNR value exceeded 10; measurements below this threshold were excluded due to significant noise interference. For individual samples, the response rate was calculated as the percentage of valid VOC measurements relative to the total measurements in the sample. Samples with a response rate ≥80% were retained for further analysis. Similarly, compound-specific response rates were determined for each VOC by calculating the proportion of valid measurements across all samples. To ensure analytical robustness, only VOCs with a response rate ≥50% were included in the validated dataset, which was designated as the “valid VOC area matrix”. Secondly, the valid VOC area matrix was then log10-transformed to address heteroscedasticity, followed by normalization to account for variations in instrument response and sample loading. These steps enabled meaningful comparison of VOC abundances across samples.

A dataset of 132 participants with malignant or benign thoracic lesions was used in this study, comprising 97 malignant and 35 benign samples. For biomarker discovery and model development, the dataset was randomly split into a discovery set (60%, n = 79; 59 malignant, 20 benign) and a testing set (40%, n = 53; 38 malignant, 15 benign). The discovery set (training set) was used for feature selection and model training, while the testing set served for independent model evaluation.

To identify VOCs differentially expressed between malignant and benign thoracic lesions, two complementary approaches were employed. First, the Wilcoxon rank-sum test was used to assess the distribution of individual VOCs across the two groups, generating corresponding p-values. Second, orthogonal partial least squares-discriminant analysis (OPLS-DA) was performed to evaluate the collective contribution of VOCs to group classification and to calculate variable importance in projection (VIP) scores (25). VOCs meeting both criteria of a p - value < 0.05 and a VIP score > 1 were selected as candidate biomarkers.

Putative biomarker identification was subsequently conducted using Agilent MassHunter Qualitative Analysis 10.0 software and the NIST 17 mass spectral library. Finally, metabolic pathway-associated VOCs reported in the literature were selected as candidate biomarkers for inclusion in diagnostic model development.

Given the complexity inherent in omics data, it is essential to identify the most suitable model for the dataset at hand. To this end, five commonly used machine learning algorithms were systematically evaluated: logistic regression (LR) (26), random forest (RF) (27), k-nearest neighbors (KNN) (28), eXtreme Gradient Boosting (XGBoost) (29), and support vector machine (SVM) (30). Among these, logistic regression algorithm demonstrated the highest robustness and effectiveness, based on its superior performance across both the discovery and testing datasets. Consequently, logistic regression model was deployed for diagnostic prediction.

To minimize overfitting, a progressive feature selection approach was employed. Biomarkers were ranked by their area under the receiver operating characteristic curve (ROC-AUC) scores. A logistic regression model was trained using 5-fold cross-validation with stratified sampling, iteratively adding one feature at a time, starting with the highest-ranked biomarker. This process continued until no further significant improvement in model performance was observed.

With the optimal feature subset identified, logistic regression hyperparameters were tuned using grid search with stratified sampling. The following hyperparameters were considered: regularization method, regularization strength, early stopping criteria, and class weights. The parameter combination that yielded the highest AUC score was selected for final model training.

The final logistic regression model, incorporating the optimized feature subset and hyperparameters, was trained on the training dataset. The model was then finalized, and a classification threshold was determined using the Youden index. Subsequently, the model’s performance was evaluated independently on the validation dataset. Performance was assessed using five metrics: F1-score, accuracy, sensitivity, specificity, and AUC, along with their respective confidence intervals. Further analyses were performed using this finalized model.

Statistical analyses were performed using Python (version 3.9.18). Continuous variables are presented as mean ± standard deviation or median [min, max], as appropriate. Categorical variables are presented as counts and percentages. The Wilcoxon rank-sum test was used to compare continuous variables between independent groups (e.g., malignant vs. benign). The Chi-square test was used to compare categorical variables. ROC analysis was performed using scikit-learn python (v1.5.1). 95% confidence intervals (95% CI) for AUC, F1-score, sensitivity, specificity, and accuracy were calculated using a binomial distribution. All statistical tests were two-sided, with a significant level of α = 0.05, unless otherwise stated.

145 participants were enrolled in this study. Exclusion criteria were applied to exclude individuals outside the age range of 18 to 80 years, those who declined participation, and those who provided invalid breath samples, resulting in a final cohort of 132 eligible participants for analysis. Among these, 77 were diagnosed with lung cancer, 13 with thymoma, 7 with esophageal cancer, and 35 had benign diseases, as confirmed by pathological results ( Figure 2 ). The demographic and clinical data of these participants are presented in Table 1 . Statistical comparisons between the case and control groups were conducted on basic demographic characteristics, including age, gender, body mass index (BMI), smoking and alcohol consumption status, and family cancer history. As detailed in Table 1 , no significant difference was observed in these factors.

Initial statistical screening using the Wilcoxon rank-sum test and OPLS-DA revealed twenty-seven VOCs that exhibited differential abundance (p < 0.05) and high VIP scores (VIP > 1) when comparing exhaled breath samples from malignant and benign groups. These candidate VOCs then underwent compound identification and further refinement to exclude those associated with drug metabolism, environmental contaminants, or unrelated to the disease pathology. This rigorous filtering process ultimately yielded a final set of 18 potentially disease-relevant VOCs ( Supplementary Table S1 ).

To identify the optimal diagnostic model for differentiating benign from malignant thoracic lesions, five machine learning algorithms including logistic regression, SVM, random forest, KNN, and XGBoost were trained using the pre-selected panel of 18 VOCs. Comparison of the models revealed that logistic regression demonstrated robust performance in both the training and validation sets, achieving AUCs of 0.85 (95% CI: 0.82, 0.89) and 0.83 (0.80, 0.89), respectively ( Figure 3A ; Supplementary Table S2 ). The DeLong test indicated that logistic regression significantly outperformed the KNN, XGBoost, and SVM models in both datasets (p < 0.05). Furthermore, when compared to the random forest model, logistic regression demonstrated superior performance in the validation set (p < 0.01). Consequently, the logistic regression model was selected for further analysis and performance evaluation.

Final feature selection was conducted using the logistic regression algorithm to optimize model performance. Analysis of the AUC as a function of the number of top features revealed diminishing returns beyond 13 features. As incorporating additional features did not substantially improve the AUC, the top 13 features were selected for model development ( Figure 3B ). These identified compounds represent a diverse range of hydrocarbons, including methyl-cyclohexane, camphene, and d-limonene, as well as oxygenated species such as butanal, 1-butanol, propanoic acid, and p-cresol. Table 2 provides a comprehensive list of these compounds and their corresponding discriminant values. Analysis of the scaled VOC peak area ( Figure 3C ) demonstrated that all 13 VOCs were present at elevated levels in the malignant group (p < 0.05). A representative chromatogram of the 13 VOCs in malignant and benign samples was shown in Figure 3D .

In the training set (n = 80), the 13-VOC model demonstrated excellent performance with an AUC of 0.86 (0.83, 0.90), an accuracy of 0.83 (0.73, 0.89), a sensitivity of 0.86 (0.76, 0.93), and a specificity of 0.71 (0.50, 0.86). In the validation set (n = 52), the 13-VOC model achieved an AUC of 0.85 (0.81, 0.90), an accuracy of 0.79 (0.66, 0.88), a sensitivity of 0.82 (0.67, 0.91), and a specificity of 0.71 (0.45, 0.88), confirming its generalizability and clinical applicability ( Figure 4A ; Supplementary Table S3 ).

To further evaluate the performance of the detection model for individual cancer types, a subgroup analysis was conducted across various malignant thoracic lesions. Thymoma (n=13) and esophageal cancer (n=7) analyses are exploratory due to limited sample size and serve as hypothesis-generating observations. In the training set, the AUCs for lung cancer, thymoma, and esophageal cancer were 0.88 (0.85, 0.90), 0.81 (0.75, 0.88), and 0.80 (0.70, 0.96), respectively. In the validation set, corresponding AUCs were 0.84 (0.80, 0.90) for lung cancer, 0.86 (0.79, 1.00) for thymoma, and 0.91 (0.83, 0.95) for esophageal cancer ( Figures 4B–D ). To further visualize the model’s performance, prediction values for each participant were plotted against their actual disease status (lung cancer/thymoma/esophageal cancer vs. benign). Using a classification threshold of 0.64, the model achieved a high accuracy in the training set, correctly identifying 87.2% (75-94%) of lung cancer, 87.5% (53-98%) of thymoma, and 75% (30-95%) of esophageal cancer cases ( Figure 4E ). In the validation set, the model maintained high accuracy, correctly classifying 80% (63-91%) of lung cancer, 80% (38-96%) of thymoma, and 100% (44-100%) of esophageal cancer cases ( Figure 4F ). Additionally, the model demonstrated good specificity for benign lesions, correctly identifying 71.4% (50-86%, 45-88%) of such cases in both the training and validation sets ( Figures 4E–F ). These findings emphasize the model’s robust performance and generalized applicability in detecting various malignant thoracic lesions. Importantly, its ability to differentiate benign lesions underscores its potential to minimize unnecessary interventions and overtreatment, supporting its use in clinical practice.

Building upon previous findings, we further investigated the model’s ability to differentiate malignant and benign pulmonary lesions. In the training set (n = 59), the 13-VOC model achieved an AUC of 0.82 (0.68, 0.95), sensitivity of 0.89 (0.77, 0.95), and specificity of 0.58 (0.32, 0.81). In the validation set (n = 38), the model exhibited an AUC of 0.79 (0.57, 0.98), sensitivity of 0.80 (0.63, 0.91), and specificity of 0.63 (0.31, 0.86) ( Figure 4G ; Supplementary Table S3 ).

Early detection of lung cancer is critical in clinical practice, allowing for timely interventions and curative resections that substantially increase patient survival rates. To assess our model’s efficacy in diagnosing early lung cancer, we used the model to differentiate between various lung cancer stages and benign nodules. The predictive performance of the model was graphically demonstrated by plotting individual participant predictions against their corresponding ground truth classifications (lung cancer stages or benign nodule). With a predetermined classification cut-off at 0.64, the 13-VOC model demonstrated strong performance in identifying early-stage lung cancer, achieving high accuracy for stage 0 + I + II lung cancer (85.7% [70.6-93.7%]) and stage III + IV lung cancer (88.9% [56.5-98%]) in the training set, though the accuracy for benign nodules was comparatively lower at 58.3% (32-80.7%) ( Figure 4H ). In the validation set, the model maintained robust performance for stage 0 + I + II lung cancer (81.8% [61.5-92.7%]) and improved accuracy for benign nodules (66.7% [35.4-87.9%]), though there was a slight decrease in accuracy for stage III+IV lung cancer (71.4% [35.9-91.8%]) ( Figure 4I ). These findings highlight the model’s potential for early detection and timely treatment of lung cancer, as well as its capacity to reduce unnecessary interventions and overtreatment—an essential consideration in clinical decision-making. However, further optimization is necessary to enhance its ability to accurately differentiate benign nodules and address variability in diagnosing advanced-stage lung cancer.

To determine whether the model represents an advancement in tumor diagnosis, we compared its predictive accuracy against that of four established clinical tumor biomarkers: CA125, ProGRP, CEA, and CFRA21-1. Among the 36 lung cancers patients, the discriminative sensitivities of CA125, ProGRP, CEA, and CFRA21–1 were 0.061, 0.121, 0.152 and 0.242 respectively, while the 13-VOCs model showed a paired discriminative sensitivity of 0.895 (p < 0.001) ( Figures 5A–D, F ). Given the common clinical practice of combining these four biomarkers to enhance specificity, we hypothesized that classifying individuals as positive for lung cancer if any of the serum tumor biomarkers fell outside the normal range (i.e., CA125: 0–35 KU/L, ProGRP: 0–46 ng/L, CEA: 0-5μg/L, CFRA21-1: 0–3 ng/L) could yield improved sensitivity. Of particular note, our 13-VOCs model significantly outperformed the 4-serum tumor marker panel, achieving a sensitivity of 0.895 compared to 0.394 (p < 0.001) ( Figures 5E, F ). Importantly, this superior performance was not attributable to an elevated false positive rate ( Figure 4 ). These findings suggested that the 13-VOCs model represents a more robust diagnostic tool, potentially offering particular advantages for early detection, an area where traditional serum biomarkers have limited utility.

To assess whether the model accurately reflects dynamic changes in disease status and further validate that the features it captures are closely associated with disease activity or burden, we analyzed and compared the model’s score changes between preoperative assessments and postoperative timepoints, 7 days to 1 month after surgery. Among the cancer patients (n = 54), postoperative predicted probabilities were significantly lower than preoperative probabilities (p < 0.01), indicating a measurable decrease in predicted disease burden following surgical intervention ( Figure 6A ). Subgroup analysis confirmed this trend in both lung cancer (p < 0.05) ( Figure 6C ) and thymoma (p < 0.05) ( Figure 6D ), with postoperative scores remaining consistently lower across these malignancies. The postoperative reduction in esophageal cancer was not statistically significant (p > 0.05) ( Figure 6E ), possibly due to the limited sample size. Notably, there was no significant difference in predicted probabilities between the postoperative and preoperative groups in cases of benign disease (p > 0.05) ( Figure 6B ). Collectively, these findings demonstrate that the model effectively reflects the reduction in disease burden following lung cancer surgery, highlighting its potential utility for assessing the completeness of resection and detecting early signs of postoperative recurrence.