Details of obtaining basic parameters such as BMI, blood pressure, collecting blood samples, and questionnaires could be referred to our previous publication.

All volunteer participants remained in a designated room under resting conditions for at least 10 min prior to sample collection. To minimize contamination, each participant was asked to rinse their mouth with unchlorinated water before exhaling through a mouthpiece connected to a three-way direct-connect valve.

Initially, exhaled breath passed through the first outlet, which was connected to a gas bag (SKC Inc., Eighty-Four, PA, United States) to estimate the volume of exhaled air. Once the volume of the initial exhalation reached approximately 0.3 L, the valve was switched to the second outlet, which was attached to a 1.0-L aluminum bag. This second bag was used to collect the late expiratory fraction, which is more representative of alveolar air and thus suitable for volatile organic compound (VOC) analysis.

To ensure adequate sample volume for analysis, the collection procedure was repeated two to three times as necessary. All collected breath samples were sealed, stored at room temperature (25°C), and analyzed within 48 h.

To validate the stability of VOCs under these storage conditions, a time-dependent analysis was conducted on ten breath samples. Samples were analyzed twice daily over three consecutive days. Comparison of the quantitative VOC data indicated that the majority of compounds remained stable during the storage period.

While numerous studies have explored the application of VOC measurements for disease identification (Ratiu et al., 2020; Keogh and Riches, 2022; Chung et al., 2022), relatively few have focused on utilizing Mach-L techniques specifically for VOC profiling (Tsou et al., 2021; Shaffie et al., 2022; Sukaram et al., 2023). In this study, we employed ten distinct Mach-L algorithms to construct predictive models for diagnosing non-alcoholic fatty liver disease (NAFLD) based on VOCs collected from exhaled breath. To assess the impact of VOCs, models were developed both with and without VOC data, and their predictive performances were compared.

The ten machine learning techniques applied are as follows:

•Random Forest (RF): An ensemble learning method utilizing multiple unpruned decision trees for classification (Breiman, 2001).

Although Mach-L algorithms are capable of identifying key predictor variables, relying on a single method may lead to suboptimal and biased feature selection. To overcome this limitation, variable ensemble strategies are often employed, which integrate the outputs from multiple algorithms. Prior research indicates that such ensemble approaches enhance variable selection robustness, reducing both bias and variance (Pes, 2020; Moghimi et al., 2018; Tuli et al., 2019).

In this study, the variable importance values generated by each Mach-L model were averaged. The top 10 VOCs, ranked by average importance across all models, were selected for further discussion.

All analyses were conducted using the R programming language (version 4.1.2, R Core Team, Vienna, Austria) and RStudio (version 1.1.453) (R Core Team, 2017; RStudio Team, 2015). The following R packages were employed for model development: random Forest, C50, gbm, RWeka, kernlab, earth, rpart, glmnet, XGBoost, LightGBM, and cat boost. Heatmaps were visualized using the pheatmap package (version 2.6.2) (Browne, 2000; Gu, 2022).

To train and evaluate each Mach-L model, an 80/20 train-test split was used. The training set (80%) was used to construct models, while the testing set (20%) evaluated predictive performance. Hyperparameter tuning was conducted using 10-fold cross-validation (CV) to ensure optimal performance for each algorithm. The final model for each method was selected based on the best-performing configuration. Cross-validation procedures were executed using the caret package (version 6.0-93) (Kuhn, 2022).

In order to understand the directions and impacts of the variables, XGboost SHAP was applied using the following Python packages: SHAP, the core package for computing and visualizing SHAP values, provides interpretability for model predictions and feature importance. Pandas, a powerful library for data manipulation and preprocessing, was used to manage datasets, clean data, and prepare inputs for SHAP analysis. NumPy, a fundamental package for numerical computations, supported array operations and numerical calculations required by SHAP. Matplotlib, a plotting library for creating static, interactive, and animated visualizations, was employed to generate SHAP plots, including summary plots, bar plots, and waterfall plots.te feature contributions to specific predictions.

To comprehensively evaluate the predictive performance of the Mach-L, we employed a range of widely accepted performance measures, as recommended in previous studies (Dias Canedo and Cordeiro Mendes, 2020; Hussain et al., 2021; Tomer and Sharma, 2022). Specifically, the following metrics were utilized in our analysis: accuracy (ACC), sensitivity (Sens), and specificity (Spec). These metrics provide an overall understanding of the model’s classification capabilities.

However, when dealing with imbalanced datasets, traditional metrics such as ACC, Sens, and Spec can be misleading, as they tend to be disproportionately influenced by the majority class distribution. To mitigate this issue, we additionally calculated balanced accuracy (BA) and area under the receiver operating characteristic curve (AUC)—both of which are considered more robust and reliable indicators for evaluating model performance under class imbalance conditions (di Biase et al., 2020; Hashim et al., 2021).

•Balanced Accuracy (BA) accounts for imbalanced data by averaging sensitivity and specificity.

The definitions and formulas for all performance metrics used in this study are detailed in (Tharwat, 2021).

To assess the impact of volatile organic compounds (VOCs) on model performance, we compared each Mach-L model’s predictive ability with and without VOC features. We applied DeLong’s test for pairwise comparison of AUC values between these two scenarios across all models (DeLong et al., 1988), allowing for a statistically grounded evaluation of VOCs’ contribution to predictive improvement.

The predictive performance of all 10 machine learning (Mach-L) methods is summarized in Table 3. Across all methods, Model 2—which incorporated volatile organic compounds (VOCs)—demonstrated superior performance compared to Model 1, which only included demographic, biochemical, and lifestyle variables. Specifically, accuracy (ACC), sensitivity (Sens), specificity (Spec), BA, and AUC were all improved in Model 2.

These findings suggest that the inclusion of VOCs significantly enhanced the predictive accuracy of the models in identifying individuals with NAFLD. The confusion matrices for Models 1 and 2 are presented in Figure 2, while Figure 3 illustrates the respective AUC curves for each model. Additionally, the heatmap of the top 10 VOCs identified across the Mach-L algorithms is shown in Figure 4, highlighting their relative importance in the classification task.

Table 4 presents the pairwise comparisons of AUC values for the 10 Mach-L methods, evaluating the improvement in predictive performance with the inclusion of VOCs compared to models without VOCs. The results indicate that for all methods, the inclusion of VOCs led to a statistically significant improvement in model performance, as evidenced by p-values less than 0.05 across all comparisons. These findings suggest that incorporating VOC data into the Mach-L models for NAFLD diagnosis results in significantly enhanced predictive accuracy compared to models that exclude VOCs.

Table 5 displays the most important predictive factors identified by the Mach-L methods, encompassing demographic, biochemical, lifestyle, and VOC-related variables. In total, 25 factors were selected, including the top 10 VOCs. Among the non-VOC variables, BMI emerged as the most influential predictor, followed by triglycerides (TG), uric acid (UA), fasting plasma glucose (FPG), γ -glutamyl transferase (γ -GT), gender, GPT, LDL-cholesterol, sleep duration, albumin, total bilirubin, alkaline phosphatase, GOT, HDL-cholesterol, and diastolic blood pressure (DBP). Notably, beginning from the 10th rank in overall importance, 2-propanol was the first VOC to appear. The complete list of VOCs identified is detailed in the Methods section.

The Bee Swarm plot derived from the XGBoost SHAP was shown in Figure 5. From top to the bottom listed the features selected and the higher horizontal feature indicates it is more important. Each circle represents a participant’s value impact of that feature. The red color has stronger impact whilst the blue one has less. Thus, this figure shows the direction of impact of each participants. Finally, Figure 6 shows the absolute strengths each feature from the highest to the lowest.

Lu et al. demonstrated that subchronic exposure to 2-propanol in mice induced NAFLD through dysregulation of the AMPK signaling pathway (Lu et al., 2015). Interestingly, in our study, 2- propanol levels were lower in NAFLD subjects, suggesting possible upregulation of AMPK as a protective or compensatory modality (Fang et al., 2022). However, the specific mechanistic studies on 2-propanol and fatty liver are limited, its known hepatotoxicity and the findings from animal studies support the possibility that 2-propanol exposure can contribute to fatty liver development, especially with high or prolonged exposure (Satapathy et al., 2015; World HealthOrganizatiom-INTERNATIONAL P ROGRAMME ON CHEMICALSAFETY, 1990).

Solga et al. previously reported that breath acetone was associated with NAFLD in morbidly obese patients undergoing bariatric surgery (Solga et al., 2006). This may reflect decreased d-3- hydroxybutyrate dehydrogenase activity or altered NADH levels, leading to acetone accumulation. There might be two mechanisms behind this relationship. First, enhancing ketogenesis reduces hepatic lipid accumulation in preclinical models. Exogenous ketones (e.g., β-hydroxybutyrate) show anti-inflammatory and antifibrotic effects, suggesting protective roles (Kwon et al., 2024). Second, while elevated acetone may indicate metabolic stress in early NAFLD, targeted ketone supplementation or ketogenic diets could mitigate steatosis and inflammation in specific contexts (Kwon et al., 2024). Our findings were consistent, with higher acetone levels in NAFLD subjects, though the difference was not statistically significant.

This fatty acid ester, found naturally in apricots (Prunus armeniaca), has been associated with celiac disease and IBS (pubchem, 2025). Only one prior study has investigated its link to NAFLD, finding higher prevalence in NAFLD patients (Raman et al., 2013). It might have influences on NAFLD due to the gut microbiota alterations which correlates with elevated 2-butanone (a structurally related ketone), hinting at broader metabolic disruptions involving ester-like compounds (Del et al., 2016). In the same time, methyl tert-butyl ether, another ether compound, shows epidemiological links to NAFLD risk in humans, suggesting potential shared mechanisms for ester/ether-induced metabolic dysfunction (Cui et al., 2024). Additional research is needed to elucidate its pathophysiological role.

Akesson reported that diethylethanolamine promotes the conversion of phosphatidylethanolamine to phosphatidylcholine, a hepatoprotective compound (Akesson, 1977). However, at present, there is no direct evidence linking it to NAFLD. It is well known that alcohol metabolism produces acetaldehyde and reactive oxygen species that cause fatty liver. While diethylethanolamine is not an alcohol, its metabolism might theoretically produce reactive intermediates that could similarly affect hepatic cells (Liu, 2014). Our findings—lower levels in NAFLD subjects—support this mechanism and suggest its potential protective role.

Studies in rats demonstrate that administration of carcinogenic doses of urethane leads to liver microsomal damage, including degranulation of liver microsomes, which impairs liver cell function and contributes to hepatic injury (Dani, 1983). In human, it is found that liver injury in workers exposed to N,N-dimethylformamide (DMF) (Nakasone et al., 2011; Nomiyama et al., 2001; Redlich et al., 1988), urethane levels were higher in the control group in our study. While this may indicate resistance to hepatic injury, further evidence is required to substantiate this hypothesis.

This anti-inflammatory, plant-derived compound activates CB2 receptors, reducing oxidative stress and hepatic injury in mice models (Varga et al., 2018). In the same time, it could reduce intracellular lipid accumulation, primarily by lower saturated fatty acids and modifying the lipid profile toward less harmful species (Scandiffio et al., 2023). Our findings align with this, supporting its potential therapeutic use.

Interestingly, furfural has a complex relationship with liver health. In low dose, it could improve mitochondrial function, reduced reactive oxygen species, and restoration of the NAD⁺/NADH redox balance, which is crucial for lipid metabolism and preventing fatty liver progression (Cheng et al., 2022). But when it is in high dose, a Maillard reaction product with antioxidant properties, furfural has demonstrated hepatocyte-protective effects in animal studies (Powell et al., 2014). This may explain its relevance in our NAFLD model.

There is currently no direct evidence or well-established research linking tridecane specifically to NAFLD or its progression. Tridecane is a hydrocarbon (alkane) commonly found in petroleum products and some environmental pollutants, but its direct impact on liver fat accumulation or liver metabolism has not been clearly documented. The only evidence is that tridecane has been associated with inflammation and lipid peroxidation, particularly in distinguishing NASH from non-NASH (The good scents company, 2025).

Although largely known for its use in food flavoring, this compound is a fatty acid and may reflect metabolic changes. The possible mechanisms include it is a BCFA involved in lipid metabolism; it can modulate gene expression related to fatty acid metabolism; and there is an indirect link to the NAVLD (Zhao et al., 2022; Liu et al., 2018; Pooya et al., 2012). We observed higher levels in non-NAFLD subjects, potentially due to better hepatic metabolism in healthier individuals (Yamaguchi et al., 2007).

Produced by Saccharomyces cerevisiae, this compound has antioxidant properties (Wilson et al., 2022; Agarwal et al., 2020; Landolfo et al., 2008), which may be linked to liver protection. In the same time, it is a fatty alcohol lipid molecule involved in metabolic pathways related to fatty acid and alcohol metabolism. It can be oxidized to 2-methylbutyrate, which then enters beta-oxidation to produce acetyl-CoA and propionyl-CoA, key intermediates in energy metabolism (Thompson et al., 2020). This indicates that (S)-2-methyl-1-butanol is metabolically linked to fatty acid catabolism through its conversion to fatty acid derivatives that feed into mitochondrial energy pathways. While no prior study has evaluated its relationship with NAFLD, our findings suggest it may play a role in hepatic defense mechanisms.

Despite the strengths of our study—including a large sample size and comprehensive VOC profiling—there are important limitations. In the present study, we do have limitations. First, this is a cross-sectional study which is less persuasive than a longitudinal one. There is no conclusion of cause-effect relationship could be drawn. However, since some of these participants will continue to have a health check up in our clinic, in the future, we believe that we will have a longitudinal study. Second, the participants were only limited to Taiwanese. It should be exercised to another ethnic group with cautious. In the future, since these participants will remain to be followed up in the MJ clinic, longitudinal studies could be done by using the present results of VOCs to predict future diseases. In the same time, we will separate these participants into two groups; one for selecting the VOCs, the rest will be treated as a validation group.