Fecal volatile organic compound–based machine learning model for noninvasive detection of colorectal cancer

Fecal volatile organic compound–based machine learning model for noninvasive detection of colorectal cancer | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Fecal volatile organic compound–based machine learning model for noninvasive detection of colorectal cancer Qi-Jun Li, Zu-Bai Li, Bo-Rong Yu, Xiao-Hong Wang, Xiao-Wen Li This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8505402/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract BACKGROUND Colorectal cancer (CRC) remains a major global health concern, ranking among the top causes of cancer incidence and mortality. Current noninvasive screening tools such as fecal occult blood tests and serum carcinoembryonic antigen (CEA) assays suffer from limited sensitivity and specificity, while colonoscopy, the diagnostic gold standard, is invasive and costly. Volatile organic compounds (VOCs), metabolic end-products influenced by gut microbiota and tumor metabolism, offer a promising avenue for noninvasive CRC detection when coupled with advanced computational modeling. AIM To develop and validate a fecal VOC-based machine learning model for noninvasive CRC detection. METHODS Fecal samples from 78 CRC patients and 57 healthy controls were analyzed using gas chromatography–ion mobility spectrometry (GC–IMS). Recursive feature elimination with cross-validation (RFECV) integrating LASSO, random forest, and support vector machine identified key VOCs. Five machine learning algorithms were constructed and optimized, and their diagnostic performance, calibration, and clinical utility were evaluated. SHapley Additive exPlanations (SHAP) analysis was applied to interpret model predictions. RESULTS Among 85 identified VOCs, 11 were consistently selected as discriminative biomarkers, including 3-methylbutanoic acid-M, indole, and 1-butanol. The XGBoost model achieved the best performance with an area under the receiver operating characteristic curve (AUROC) of 0.8866, sensitivity of 0.83, and specificity of 0.78. SHAP analysis revealed 3-methylbutanoic acid-M as the most influential metabolite in model predictions. Several individual VOCs, such as 2-phenylacetaldehyde and propanoic acid-D, outperformed CEA in discriminating CRC from healthy controls. Decision curve analysis demonstrated superior clinical net benefit for the VOC-based model compared with traditional screening markers. CONCLUSION Integration of fecal VOC profiling with a machine learning model provides a promising noninvasive strategy for accurate CRC detection, potentially improving early diagnosis and screening compliance. Trial Registration Chinese Clinical Trial Registry (ChiCTR), ChiCTR2300073117. Registered on July 1, 2023 expected completion on June 30, 2025. Available at https//www.chictr.org.cn/bin/project/edit?pid=200842 Colorectal cancer Volatile organic compounds Fecal biomarkers Machine learning Noninvasive detection Gas chromatography–ion mobility spectrometry Diagnostic model SHAP analysis XGBoost Early screening Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Core Tip This study presents a novel, noninvasive diagnostic strategy for colorectal cancer (CRC) based on fecal volatile organic compound (VOC) profiling integrated with machine learning. Using GC–IMS analysis and a robust feature selection pipeline, 11 key VOC biomarkers were identified. The XGBoost model achieved high diagnostic accuracy (AUROC = 0.8866) and superior clinical utility compared with carcinoembryonic antigen. SHAP interpretation revealed 3-methylbutanoic acid-M as the most influential biomarker. This VOC-based machine learning model offers a promising, accurate, and patient-friendly approach for early CRC detection and may improve compliance with large-scale screening programs. INTRODUCTION Colorectal cancer (CRC) ranks among the most prevalent and lethal malignancies worldwide, representing a major global health challenge. According to the 2022 global cancer statistics report, CRC has the third-highest incidence and the second-highest mortality rate among all cancer(Bray et al., 2024 ). The five-year survival rate for patients diagnosed at an early stage can reach 90%, but it declines sharply once the disease advances. Commonly used screening methods include fecal occult blood testing (FOBT), colonoscopy, carcinoembryonic antigen (CEA) measurement, and radiological imaging. Although colonoscopy remains the diagnostic gold standard for CRC(Rutter et al., 2018 ), its invasive nature, high cost, and requirement for bowel preparation restrict its utility for large-scale population screening(Quintero et al., 2012 ). FOBT, despite its convenience and noninvasive characteristics, suffers from limited sensitivity and specificity. Computed tomographic colonography (CTC) provides a noninvasive alternative but is less effective at detecting small or flat lesions and exposes patients to radiation while incurring substantial expense. Similarly, CEA serves as a general tumor marker for gastrointestinal malignancies, but its diagnostic role is auxiliary, with reported sensitivity and specificity of only 53% and 86%, respectively, precluding its use as a standalone diagnostic indicator(Acharya et al., 2017 ). Volatile organic compounds (VOCs), which are abundantly present in easily obtainable biological matrices such as exhaled breath, feces, and urine, have emerged as promising candidates for CRC screening because of their noninvasive nature, convenience, and low cost(Barbosa and Filho, 2024 ; Li et al., 2024 ; Markar et al., 2015 ). A meta-analysis of 32 studies reported that VOC-based testing achieved a sensitivity of 0.88 and a specificity of 0.85 in differentiating CRC patients from healthy controls (HC), with an area under the receiver operating characteristic curve (AUROC) of 0.93(Wang et al., 2024 ), providing strong evidence for its clinical potential. However, the high-dimensional and complex nature of VOC data presents significant analytical challenges. In this regard, machine learning (ML) offers an effective solution by identifying the most discriminative feature combinations from large datasets to develop robust diagnostic model(Handelman et al., 2018 ). Furthermore, the value of ML is reflected in two major aspects: first, its ability to integrate multi-source heterogeneous data, thereby constructing more comprehensive and individualized diagnostic models; and second, its dynamic learning capacity, which enables real-time monitoring of disease progression and treatment response(Esteva et al., 2019 ). By capturing intricate, non-linear relationships within complex biological datasets, ML holds remarkable potential for early CRC detection, assessment of therapeutic efficacy, and prognosis prediction, capabilities that surpass those of traditional statistical approaches(Reel et al., 2021 ). However, the current field of VOC detection lacks standardized analytical protocols, resulting in a low concordance rate, less than 30%, of characteristic VOCs identified across different studies(Hanna et al., 2019 ), which directly undermines the comparability and reproducibility of findings. Moreover, VOC profiles are highly susceptible to external influences such as dietary habits, medication use, and environmental factors, suggesting that their diagnostic specificity still requires substantial improvement. To address the ongoing clinical challenges in early CRC diagnosis, this study aims to develop and validate an integrated diagnostic model that combines fecal VOC profiling with machine learning techniques. By optimizing both sample processing procedures and algorithmic workflows, we seek to identify key VOC biomarkers from high-dimensional data and construct a noninvasive diagnostic tool with superior accuracy compared to conventional tests. We anticipate that this approach will offer a novel, accessible, and high-performance solution to overcome the current limitations of CRC screening in sensitivity, specificity, and clinical applicability. MATERIALS AND METHODS Overview of participants and samples A total of 135 participants were enrolled in this study at Huadong Hospital, Fudan University (Shanghai, China) between July 2024 and June 2025, including 78 patients with pathologically confirmed CRC and 57 HC. All subjects were recruited from the Yangtze River Delta region, a population characterized by relatively homogeneous dietary habits, thereby minimizing dietary confounding effects. The inclusion criteria were as follows: (1) age between 25 and 80 years; (2) histopathological confirmation of adenocarcinoma for CRC cases; and (3) willingness to participate and provision of written informed consent. HC participants were required to show no evidence of CRC or other organic gastrointestinal diseases, as verified by colonoscopy and imaging, and to have no history of malignancy or infectious disease. Exclusion criteria included: (1) pregnancy, lactation, or potential for pregnancy; (2) presence of congenital disorders; (3) history of psychiatric illness; (4) active peptic ulcer; (5) acute illness within two weeks prior to enrollment; (6) history of infectious diseases; (7) prior exposure to chemotherapy, radiotherapy, or combined oncologic treatments; (8) severe chronic obstructive pulmonary disease; (9) unstable diabetes mellitus; and (10) history of other malignancies. All participants provided written informed consent before sample collection. The study protocol was reviewed and approved by the Ethics Committee of Huadong Hospital Affiliated to Fudan University (Approval No. KY 2023K127). VOCs analysis methods and procedures Approximately 0.5 g of each fecal sample was accurately weighed and transferred into a 20 mL headspace vial, followed by the addition of 1 mL of distilled water and 0.5 g of aspartic acid powder. The mixture was incubated at 60 °C for 5 min, after which 1 mL of headspace gas was extracted using a gastight syringe (Hamilton, Reno, NV, USA) and immediately injected into the analytical system, with nitrogen (99.999% purity) serving as the carrier gas. VOCs were analyzed using a FlavourSpec® gas chromatography–ion mobility spectrometry (GC–IMS) system (G.A.S., Dortmund, Germany). Within this system, VOC components were initially separated through a GC column and subsequently introduced into the IMS detector for secondary separation and detection. Each VOC was identified in two dimensions, based on its GC retention index (RI) and IMS drift time (Dt), and quantified semi-quantitatively according to signal intensity. Selection of feature VOCs To identify key fecal VOC biomarkers associated with CRC, all data were normalized and subjected to feature selection using Recursive Feature Elimination with Cross-Validation (RFECV). This process was independently conducted using three core estimators: Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and a hyperparameter-optimized Support Vector Machine (SVM). Each algorithm was tuned and validated through 5-fold cross-validation to determine its optimal subset of discriminative features. The final core VOC panel was established by intersecting the selected features shared across all three algorithms, ensuring the robustness and reproducibility of biomarker selection. Construction and evaluation of the predictive model Based on the selected common VOC biomarkers, five machine learning models were constructed: eXtreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), Adaptive Boosting (AdaBoost), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM). The complete dataset was randomly divided into training and independent testing sets in a 7:3 ratio. Hyperparameter optimization was performed on the training set using the RandomizedSearchCV function (Scikit-learn, Version 1.5.0; Python Software Foundation, Wilmington, DE, USA), sampling 20 random parameter combinations from a predefined grid. The optimization process was guided by the area under the receiver operating characteristic curve (AUROC), evaluated through 5-fold stratified cross-validation. Model performance was comprehensively assessed using multiple indicators. On the training set, the mean and standard deviation of AUROC, area under the precision-recall curve (AUPRC), and Brier score were calculated via 5-fold stratified cross-validation. Decision curve analysis (DCA) was also conducted using the cross-validated predictions to estimate clinical utility. On the independent test set, the same metrics were computed, a confusion matrix was generated, and DCA was validated. To enhance interpretability, the best-performing model on the test set was further analyzed using SHapley Additive exPlanations (SHAP; Lundberg and Lee, University of Washington, Seattle, WA, USA), and the feature contributions were visualized through a beeswarm plot. RESULTS Study design and baseline characteristics of the study participants A total of 78 CRC patients and 57 HC were included in this study. Fecal samples were collected from all participants, and VOC data were obtained using GC–IMS. Machine learning algorithms were subsequently applied to identify informative VOCs and construct a diagnostic model for CRC detection (Figure 1A). Baseline demographic analysis demonstrated no statistically significant differences in age, sex, or body mass index (BMI) between the CRC and HC groups ( P > 0.05) (Table S1). As illustrated in Figure 1B, the proportions of male and female participants were comparable, indicating balanced gender distribution within the cohort. The age distribution profiles (violin bodies) of the two groups are highly overlapping, suggesting that the study population was well matched in terms of age structure. Analysis of fecal volatile organic compounds A total of 117 VOCs were detected in fecal samples. After excluding 32 compounds that could not be confidently identified, 85 VOCs were retained for subsequent analysis. Following the Level 1 identification criteria established by the Metabolomics Standards Initiative, these VOCs were classified into major chemical categories associated with CRC. The overall chemical composition consisted of alcohols (20%), fatty acids (17.65%), aldehydes (14.12%), ketones and esters (10.59% each), sulfur-containing compounds (5.88%), and other compounds (21.18%), including alkaloids, alkanes, furans, terpenes, and pyrazines (Figure 1C, Table S2). This classification reflects the chemical diversity of fecal VOCs and highlights multiple metabolite classes potentially implicated in CRC-related metabolic alterations. Screening results of feature VOCs The feature selection performance curves for RFE-LASSO, RFE-RF, and RFE-SVM are presented in Figure 2A–C (Table S3). A Venn diagram (Figure 2D) illustrates the overlap among the VOC features identified by the three algorithms, highlighting 11 common VOCs consistently selected as key discriminatory biomarkers: (E)-2-pentenal, nonanal-D, 2-phenylacetaldehyde, indole, dimethyl trisulfide, 1-butanol, 4-methylphenol, pentanoic acid-M, propanoic acid-D, nonanal-M, and 3-methylbutanoic acid-M. These shared metabolites represent the most robust features contributing to CRC discrimination across multiple machine learning frameworks. VOCs-based XGBoost model for accurate diagnosis and interpretability analysis To comprehensively evaluate model performance, receiver operating characteristic (ROC) curves, precision–recall (PR) curves, and decision curve analyses (DCA) were generated for both the training and test sets, along with comparative assessments of Brier scores. During cross-validation on the training set (Figure 3A–B), all models achieved ROC curves that were substantially higher than the random classification baseline. Among them, CatBoost demonstrated the highest AUROC, followed by LightGBM, XGBoost, RF, and AdaBoost in descending order. On the independent test set, all models maintained superior classification performance compared with random guessing. Notably, XGBoost exhibited excellent generalization ability, with a test AUROC nearly identical to its training result. At the optimal threshold, XGBoost achieved a sensitivity of 0.83 and a specificity of 0.78, reflecting a strong balance between detection and false-positive rates. In the PR analysis of the training set (Figure 3C), XGBoost achieved an AUPRC of 0.9121, ranking second only to CatBoost. However, in the test set (Figure 3D), its performance improved further, reaching an AUPRC of 0.9251, the highest among all evaluated models. DCA revealed that all models offered a substantially greater net clinical benefit than the “treat-all” and “treat-none” strategies across the 0–0.8 probability threshold range (Figure 3E). On the test set, XGBoost demonstrated the highest net benefit within the high-threshold probability range (0.8–1.0) (Figure 3F). Comparison of Brier scores (Figure 3G) showed minimal differences between the training and test sets across all models, with XGBoost achieving the lowest score, indicating superior calibration. The SHAP feature importance analysis for XGBoost (Figure 4A) identified 3-methylbutanoic acid-M as the most influential feature, exhibiting the broadest distribution and largest absolute SHAP values, followed by indole and 1-butanol. The SHAP dependence plot (Figure 4B) further demonstrated a positive relationship between the concentration of 3-methylbutanoic acid-M and the probability of CRC prediction, underscoring its critical contribution to model decision-making. Analysis of diagnostic efficacy of biomarkers and expression differences of VOCs between CRC and HC Compared with the clinically established tumor marker CEA (Table S4), which achieved an AUROC of 0.668 (Figure 5A), a value slightly higher than previously reported, its standalone diagnostic value for CRC remained limited. In contrast, several individual VOCs, including indole, 2-phenylacetaldehyde, 3-methylbutanoic acid-M, propanoic acid-D, and 1-butanol, demonstrated higher AUROC values than CEA (Figure 5B), indicating their superior diagnostic potential. As shown in Figure 5C, the concentrations of these metabolites differed significantly between groups. CRC patients exhibited markedly reduced levels of 1-butanol, 3-methylbutanoic acid-M, and propanoic acid-D compared with HC ( P < 0.001), whereas the level of 2-phenylacetaldehyde was significantly elevated ( P < 0.001). These findings highlight distinct metabolic alterations in fecal VOC profiles associated with CRC and underscore their potential as noninvasive biomarkers for early detection. DISCUSSION This study aimed to establish a noninvasive diagnostic strategy for CRC using fecal metabolomic profiling integrated with machine learning. From 78 CRC and 57 HC samples, a total of 85 VOCs were identified through advanced analytical techniques. Employing an RFECV-based feature selection framework allowed the systematic extraction of robust biomarkers, effectively minimizing feature redundancy and selection bias often reported in earlier studies(Al-Tashi et al., 2023 ; Zhang and Liu, 2021 ). Among the models tested, XGBoost demonstrated the most consistent and superior performance across multiple evaluation metrics, particularly those relevant to clinical diagnostic accuracy. At the optimal decision threshold, the model achieved a sensitivity of 0.83 and a specificity of 0.78, meeting the clinical demands for both reliable detection and minimal misclassification(van Doorn et al., 2021 ). The strong performance of XGBoost can be attributed to its gradient boosting structure, which iteratively refines residuals and dynamically adjusts feature weights, thereby capturing intricate non-linear interactions between VOCs and disease state. This finding is consistent with previous evidence highlighting the capacity of gradient boosting algorithms to handle high-dimensional biomarker data with exceptional precision(Guo et al., 2025 ). Furthermore, SHAP analysis provided interpretability by elucidating the individual and combined effects of key VOCs on model predictions, thereby addressing the “black box” limitation of conventional machine learning models and offering mechanistic insight into the biological relevance of predictive features(Rudin, 2019 ). We identified 11 key VOCs, comprising three short-chain fatty acids (SCFAs), one sulfur-containing compound, four aldehydes, two alcohols, and indole. Comparative analysis against the widely recognized CRC screening biomarker CEA demonstrated that 1-butanol, 2-phenylacetaldehyde, 3-methylbutanoic acid-M, and propanoic acid-D exhibited significantly superior discriminatory power for CRC detection, underscoring their potential utility as dynamic indicators for disease monitoring. The reduction in SCFAs and their derivatives, including propanoic acid-D, pentanoic acid-M, and 3-methylbutanoic acid-M, was consistent with findings from previous studies(Ou et al., 2013 ; Weir et al., 2013 ). Lower SCFA concentrations may compromise intestinal barrier integrity and immune homeostasis, thereby facilitating CRC development and progression(Silva et al., 2020 ). Moreover, these metabolites serve as sensitive indicators of gut microbial metabolic activity. For example, diminished levels of propanoic acid-D may reflect a decreased abundance of SCFA-producing bacterial taxa or impaired carbohydrate fermentation capacity within the gut microbiome(Cong et al., 2022 ). Aldehydes such as (E)-2-pentenal, nonanal-D, and 2-phenylacetaldehyde are likely generated as by-products of lipid peroxidation and oxidative stress. During oxidative stress, reactive oxygen species attack polyunsaturated fatty acids within cell membranes, initiating a cascade of lipid peroxidation reactions that yield diverse VOCs, including aldehydes and alcohols, which can modulate cytotoxicity and apoptosis(Boots et al., 2012 ; West and Marnett, 2006 ). Among these, 2-phenylacetaldehyde, a crucial intermediate in phenylalanine metabolism, may participate in the biosynthesis of aromatic compounds and contribute to the regulation of cellular signaling pathways(Debnar-Daumler et al., 2014 ; Tieman et al., 2006 ). Sulfur-containing compounds, such as dimethyl trisulfide, are primarily produced by gut microbiota through the degradation of dietary sulfur-containing amino acids and are closely linked to intestinal inflammation and oxidative stress(Kumar et al., 2025 ). The observed alterations in fecal VOCs were consistent with changes previously reported in breath VOC profiles from our earlier study(Liu et al., 2024 ), suggesting a potential metabolic link between fecal and breath volatile signatures. This concordance supports the hypothesis that systemic and intestinal metabolic processes contribute jointly to CRC-associated volatilomic changes. Furthermore, accumulating evidence indicates that gut microbiota may influence CRC pathogenesis through multiple mechanisms, including the production of bioactive metabolites, induction of chronic inflammation, and modulation of cancer cell energy metabolism(Yang et al., 2019 ). These insights provide a valuable framework for future mechanistic investigations into the microbiota–metabolite–cancer axis. This study, however, has several limitations. The cohort was derived from a single center with a modest sample size, which may introduce selection bias and limit the generalizability of the findings. All samples were collected at a single time point, precluding assessment of longitudinal changes in VOCs during disease progression. Due to the limited cohort size, subgroup analyses based on CRC stage or molecular subtype could not be performed.Despite these limitations, this study introduces a promising and practical framework for noninvasive CRC diagnosis by integrating VOC profiling with machine learning-based modeling. To address these limitations, we propose three strategies for subsequent research. First, we have established a collaboration with one clinical center in Southwest China and plan to collect at least 400 samples. By comparing model performance metrics such as AUC and sensitivity across data from different centers, we will systematically evaluate the impact of external factors, further refine the model, and establish a quantitative correction framework for confounding variables. Second, we will improve sample collection protocols by conducting comparative validation of sampling at different time points within the same day, aiming to achieve end-to-end standardization from sample collection to detection. Third, we will supplement the analysis with CRC subtype stratification(Altomare et al., 2013 ) to clarify the model’s applicability boundaries across populations with different clinical stages and pathological types, thereby supporting its precise clinical translation. CONCLUSION This study developed and validated a noninvasive diagnostic approach for CRC by integrating fecal VOC profiling with advanced machine learning modeling. Using GC–IMS and a robust feature selection strategy, 11 key VOC biomarkers were identified, several of which demonstrated superior diagnostic performance compared with CEA. The XGBoost model exhibited strong discriminative power, calibration, and clinical utility, achieving a balanced sensitivity and specificity suitable for early CRC detection. Moreover, SHAP analysis provided mechanistic interpretability, revealing that specific VOCs, such as 3-methylbutanoic acid-M and indole, play dominant roles in disease classification. These findings highlight the potential of VOC-based machine learning models as accessible, accurate, and patient-friendly tools for CRC screening. Future work should focus on large-scale, multi-center validation and the development of portable platforms to enable real-world clinical application. Declarations Ethics approval and consent to participate All participants provided written informed consent. The study was approved by the Ethics Committee of Huadong Hospital Affiliated to Fudan University (Approval No. 2023K127) and was registered in the Chinese Clinical Trial Registry (ChiCTR2300073117). Consent for publication Not applicable. Acknowledgments Not applicable. Availability of data and materials The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request. Competing interests The authors declare that they have no conflict of interests Author contributions: Li XW and Li QJ conceived the study. Li ZB developed the methodology and created the visualizations. Wang XH and Yu BR contributed to data collection, Li QJ and Li ZB drafted the original manuscript. Li QJ prepared figures 1-3. Li XW prepared figures 4-5 and interpreted results. Li XW and XH.W. provided supervision throughout the project. All authors reviewed and edited the manuscript. Supported by ： This work was supported by the by the Key Specialized Disease Project (2022, ZDZB2221) and Key department (2022, ZDXK2213) of Huadong Hospital. References Acharya, A., Markar, S.R., Matar, M., Ni, M. and Hanna, G.B. (2017) Use of Tumor Markers in Gastrointestinal Cancers: Surgeon Perceptions and Cost-Benefit Trade-Off Analysis. Ann Surg Oncol 24, 1165-1173.https://doi.org/10.1245/s10434-016-5717-y Al-Tashi, Q., Saad, M.B., Muneer, A., Qureshi, R., Mirjalili, S., Sheshadri, A., Le, X., Vokes, N.I., Zhang, J. and Wu, J. (2023) Machine Learning Models for the Identification of Prognostic and Predictive Cancer Biomarkers: A Systematic Review. Int J Mol Sci 24 .https://doi.org/10.3390/ijms24097781 Altomare, D.F., Di Lena, M., Porcelli, F., Trizio, L., Travaglio, E., Tutino, M., Dragonieri, S., Memeo, V. and de Gennaro, G. (2013) Exhaled volatile organic compounds identify patients with colorectal cancer. Br J Surg 100, 144-50.https://doi.org/10.1002/bjs.8942 Barbosa, J.M.G. and Filho, N.R.A. (2024) The human volatilome meets cancer diagnostics: past, present, and future of noninvasive applications. Metabolomics 20, 113.https://doi.org/10.1007/s11306-024-02180-5 Boots, A.W., van Berkel, J.J., Dallinga, J.W., Smolinska, A., Wouters, E.F. and van Schooten, F.J. (2012) The versatile use of exhaled volatile organic compounds in human health and disease. J Breath Res 6, 027108.https://doi.org/10.1088/1752-7155/6/2/027108 Bray, F., Laversanne, M., Sung, H., Ferlay, J., Siegel, R.L., Soerjomataram, I. and Jemal, A. (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 74, 229-263.https://doi.org/10.3322/caac.21834 Cong, J., Zhou, P. and Zhang, R. (2022) Intestinal Microbiota-Derived Short Chain Fatty Acids in Host Health and Disease. Nutrients 14 .https://doi.org/10.3390/nu14091977 Debnar-Daumler, C., Seubert, A., Schmitt, G. and Heider, J. (2014) Simultaneous involvement of a tungsten-containing aldehyde:ferredoxin oxidoreductase and a phenylacetaldehyde dehydrogenase in anaerobic phenylalanine metabolism. J Bacteriol 196, 483-92.https://doi.org/10.1128/jb.00980-13 Esteva, A., Robicquet, A., Ramsundar, B., Kuleshov, V., DePristo, M., Chou, K., Cui, C., Corrado, G., Thrun, S. and Dean, J. (2019) A guide to deep learning in healthcare. Nat Med 25, 24-29.https://doi.org/10.1038/s41591-018-0316-z Guo, C., Liu, Z., Fan, H., Wang, H., Zhang, X., Zhao, S., Li, Y., Han, X., Wang, T., Chen, X. and Zhang, T. (2025) Machine-learning-based plasma metabolomic profiles for predicting long-term complications of cirrhosis. Hepatology 81, 168-180.https://doi.org/10.1097/hep.0000000000000879 Handelman, G.S., Kok, H.K., Chandra, R.V., Razavi, A.H., Lee, M.J. and Asadi, H. (2018) eDoctor: machine learning and the future of medicine. J Intern Med 284, 603-619.https://doi.org/10.1111/joim.12822 Hanna, G.B., Boshier, P.R., Markar, S.R. and Romano, A. (2019) Accuracy and Methodologic Challenges of Volatile Organic Compound-Based Exhaled Breath Tests for Cancer Diagnosis: A Systematic Review and Meta-analysis. JAMA Oncol 5, e182815.https://doi.org/10.1001/jamaoncol.2018.2815 Kumar, R., Sykes, D.J., Band, V.I., Schaller, M.L., Patel, R., Vitvitsky, V., Sajjakulnukit, P., Singhal, R., Wong, H.K.A., Hourigan, S.K., Ichinose, F., Lyssiotis, C.A., Shah, Y.M. and Banerjee, R. (2025) Gut sulfide metabolism modulates behavior and brain bioenergetics. Proc Natl Acad Sci U S A 122, e2503677122.https://doi.org/10.1073/pnas.2503677122 Li, X., Shi, L., Long, Y., Wang, C., Qian, C., Li, W., Tian, Y. and Duan, Y. (2024) Volatile organic compounds in exhaled breath: a promising approach for accurate differentiation of lung adenocarcinoma and squamous cell carcinoma. J Breath Res 18 .https://doi.org/10.1088/1752-7163/ad6474 Liu, Y., Ji, Y., Chen, J., Zhang, Y., Li, X. and Li, X. (2024) Pioneering noninvasive colorectal cancer detection with an AI-enhanced breath volatilomics platform. Theranostics 14, 4240-4255.https://doi.org/10.7150/thno.94950 Markar, S.R., Wiggins, T., Kumar, S. and Hanna, G.B. (2015) Exhaled breath analysis for the diagnosis and assessment of endoluminal gastrointestinal diseases. J Clin Gastroenterol 49, 1-8.https://doi.org/10.1097/mcg.0000000000000247 Ou, J., Carbonero, F., Zoetendal, E.G., DeLany, J.P., Wang, M., Newton, K., Gaskins, H.R. and O'Keefe, S.J. (2013) Diet, microbiota, and microbial metabolites in colon cancer risk in rural Africans and African Americans. Am J Clin Nutr 98, 111-20.https://doi.org/10.3945/ajcn.112.056689 Quintero, E., Castells, A., Bujanda, L., Cubiella, J., Salas, D., Lanas, Á., Andreu, M., Carballo, F., Morillas, J.D., Hernández, C., Jover, R., Montalvo, I., Arenas, J., Laredo, E., Hernández, V., Iglesias, F., Cid, E., Zubizarreta, R., Sala, T., Ponce, M., Andrés, M., Teruel, G., Peris, A., Roncales, M.P., Polo-Tomás, M., Bessa, X., Ferrer-Armengou, O., Grau, J., Serradesanferm, A., Ono, A., Cruzado, J., Pérez-Riquelme, F., Alonso-Abreu, I., de la Vega-Prieto, M., Reyes-Melian, J.M., Cacho, G., Díaz-Tasende, J., Herreros-de-Tejada, A., Poves, C., Santander, C. and González-Navarro, A. (2012) Colonoscopy versus fecal immunochemical testing in colorectal-cancer screening. N Engl J Med 366, 697-706.https://doi.org/10.1056/NEJMoa1108895 Reel, P.S., Reel, S., Pearson, E., Trucco, E. and Jefferson, E. (2021) Using machine learning approaches for multi-omics data analysis: A review. Biotechnol Adv 49, 107739.https://doi.org/10.1016/j.biotechadv.2021.107739 Rudin, C. (2019) Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead. Nat Mach Intell 1, 206-215.https://doi.org/10.1038/s42256-019-0048-x Rutter, M.D., Beintaris, I., Valori, R., Chiu, H.M., Corley, D.A., Cuatrecasas, M., Dekker, E., Forsberg, A., Gore-Booth, J., Haug, U., Kaminski, M.F., Matsuda, T., Meijer, G.A., Morris, E., Plumb, A.A., Rabeneck, L., Robertson, D.J., Schoen, R.E., Singh, H., Tinmouth, J., Young, G.P. and Sanduleanu, S. (2018) World Endoscopy Organization Consensus Statements on Post-Colonoscopy and Post-Imaging Colorectal Cancer. Gastroenterology 155, 909-925.e3.https://doi.org/10.1053/j.gastro.2018.05.038 Silva, Y.P., Bernardi, A. and Frozza, R.L. (2020) The Role of Short-Chain Fatty Acids From Gut Microbiota in Gut-Brain Communication. Front Endocrinol (Lausanne) 11, 25.https://doi.org/10.3389/fendo.2020.00025 Tieman, D., Taylor, M., Schauer, N., Fernie, A.R., Hanson, A.D. and Klee, H.J. (2006) Tomato aromatic amino acid decarboxylases participate in synthesis of the flavor volatiles 2-phenylethanol and 2-phenylacetaldehyde. Proc Natl Acad Sci U S A 103, 8287-92.https://doi.org/10.1073/pnas.0602469103 van Doorn, W., Stassen, P.M., Borggreve, H.F., Schalkwijk, M.J., Stoffers, J., Bekers, O. and Meex, S.J.R. (2021) A comparison of machine learning models versus clinical evaluation for mortality prediction in patients with sepsis. PLoS One 16, e0245157.https://doi.org/10.1371/journal.pone.0245157 Wang, Q., Fang, Y., Tan, S., Li, Z., Zheng, R., Ren, Y., Jiang, Y. and Huang, X. (2024) Diagnostic performance of volatile organic compounds analysis and electronic noses for detecting colorectal cancer: a systematic review and meta-analysis. Front Oncol 14, 1397259.https://doi.org/10.3389/fonc.2024.1397259 Weir, T.L., Manter, D.K., Sheflin, A.M., Barnett, B.A., Heuberger, A.L. and Ryan, E.P. (2013) Stool microbiome and metabolome differences between colorectal cancer patients and healthy adults. PLoS One 8, e70803.https://doi.org/10.1371/journal.pone.0070803 West, J.D. and Marnett, L.J. (2006) Endogenous reactive intermediates as modulators of cell signaling and cell death. Chem Res Toxicol 19, 173-94.https://doi.org/10.1021/tx050321u Yang, Y., Misra, B.B., Liang, L., Bi, D., Weng, W., Wu, W., Cai, S., Qin, H., Goel, A., Li, X. and Ma, Y. (2019) Integrated microbiome and metabolome analysis reveals a novel interplay between commensal bacteria and metabolites in colorectal cancer. Theranostics 9, 4101-4114.https://doi.org/10.7150/thno.35186 Zhang, Z. and Liu, Z.P. (2021) Robust biomarker discovery for hepatocellular carcinoma from high-throughput data by multiple feature selection methods. BMC Med Genomics 14, 112.https://doi.org/10.1186/s12920-021-00957-4 Additional Declarations No competing interests reported. Supplementary Files 11306mQACCminimumreportingstandardchecklist.pdf Supplementarymaterial.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8505402","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":571032413,"identity":"15c87fa2-10c0-429a-b6ef-25874b9cfa2f","order_by":0,"name":"Qi-Jun Li","email":"","orcid":"","institution":"Huadong Hospital Affiliated to Fudan UniversityShanghai","correspondingAuthor":false,"prefix":"","firstName":"Qi-Jun","middleName":"","lastName":"Li","suffix":""},{"id":571032415,"identity":"da0b8cfb-1885-4f84-a6cc-9341c0477722","order_by":1,"name":"Zu-Bai Li","email":"","orcid":"","institution":"Shanghai University of Engineering Sciences","correspondingAuthor":false,"prefix":"","firstName":"Zu-Bai","middleName":"","lastName":"Li","suffix":""},{"id":571032417,"identity":"29c0efb5-a9a8-4896-8211-d4354f3717b2","order_by":2,"name":"Bo-Rong Yu","email":"","orcid":"","institution":"Huadong Hospital Affiliated to Fudan UniversityShanghai","correspondingAuthor":false,"prefix":"","firstName":"Bo-Rong","middleName":"","lastName":"Yu","suffix":""},{"id":571032418,"identity":"ba589be2-de4e-481d-93e6-3f18673ea6dc","order_by":3,"name":"Xiao-Hong Wang","email":"","orcid":"","institution":"Huadong Hospital Affiliated to Fudan UniversityShanghai","correspondingAuthor":false,"prefix":"","firstName":"Xiao-Hong","middleName":"","lastName":"Wang","suffix":""},{"id":571032421,"identity":"fe3e77ec-d470-4b77-b74c-254d0782c91d","order_by":4,"name":"Xiao-Wen Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/0lEQVRIiWNgGAWjYDACZgglw8bMfPBBBZQrQYwWHjZ2tmSDM0RpgQIeBn4eMwmitBgcZ374mKfiDg8fM4NZxcE2a3lzBuaDt3kY7PJwaZFsZjM25jnzjIeNmSHtxsG2dMOdDWzJ1jwMycW4tPADDZfObTsM0nLs9se2w4wbDvCYSfMwHEhswKGFjZn9m3TuP5AWxraCg22H7Tcc4P+GVws/M9DM3AaQFmY2BqCWRKAtbHi1SDbzFBv/OQbSwsYsceBcevKGw2zGlnMMknFqMTh/fOPDGTWH5eT7z3/8cKDM2nbD8eaHN95U2OHUggWAo8aAePWjYBSMglEwCjABACjqTpbIKa55AAAAAElFTkSuQmCC","orcid":"","institution":"Huadong Hospital Affiliated to Fudan UniversityShanghai","correspondingAuthor":true,"prefix":"","firstName":"Xiao-Wen","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2026-01-03 08:23:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8505402/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8505402/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100008976,"identity":"e385c280-7d7f-42a8-9c01-9b18dee25677","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1238942,"visible":true,"origin":"","legend":"","description":"","filename":"Fecalvolatileorganiccompoundbasedmachinelearningmodelfornoninvasivedetectionofcolorectalcancer.docx","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/783ae950eefae58003ddf0cf.docx"},{"id":100008984,"identity":"4fd53e0e-d1c5-400b-bae5-23590414568f","added_by":"auto","created_at":"2026-01-12 06:00:16","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7318,"visible":true,"origin":"","legend":"","description":"","filename":"d4de8742548b463287955a45e5333caf.json","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/17b1caaeaa873d8558dff7e7.json"},{"id":100008963,"identity":"b84f8882-2a9f-4768-a303-91c3c7bf88d7","added_by":"auto","created_at":"2026-01-12 06:00:14","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":178624,"visible":true,"origin":"","legend":"","description":"","filename":"11306mQACCminimumreportingstandardchecklist.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/9bc1099b2e5d5bf0cf107061.pdf"},{"id":100008975,"identity":"64008c66-a5b3-4e65-b1a1-9b9d987f3de2","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":300411,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/cdeb9744af7556226cdc3da4.pdf"},{"id":100361780,"identity":"45ad8735-1c9e-4f77-a76c-c10f18be60e8","added_by":"auto","created_at":"2026-01-16 07:45:43","extension":"xml","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":106762,"visible":true,"origin":"","legend":"","description":"","filename":"d4de8742548b463287955a45e5333caf1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/3de6dc0487109af2ce1656e9.xml"},{"id":100008989,"identity":"acab572a-eda9-4fe9-a3a4-b26fe57d0313","added_by":"auto","created_at":"2026-01-12 06:00:16","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":737794,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/dedc60928126ee4b31c48fef.jpeg"},{"id":100008971,"identity":"5513a765-db73-4f92-9171-64df99148478","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":220710,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/7b2e520838b340ef1ebfa567.jpeg"},{"id":100008977,"identity":"1877e23a-6886-47d3-b409-6740bead08da","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":702208,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/3b1258c0a601c9643691e5d7.jpeg"},{"id":100008979,"identity":"a7d819fb-c149-4362-ba51-20ad6adaf12b","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":339554,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/82d3212ecab2a7f66e42f811.jpeg"},{"id":100008983,"identity":"4c5f4b5c-613f-4708-9c22-e768fedcaa34","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"emf","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2523948,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.emf","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/5507bf1d804ec91c5e166744.emf"},{"id":100008987,"identity":"1c981feb-011f-475e-944f-865d5a7ea1a2","added_by":"auto","created_at":"2026-01-12 06:00:16","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":53466,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/7a0843947b08020a6c875006.png"},{"id":100008990,"identity":"96ce8193-b8d4-46e4-bc14-a19935ee876e","added_by":"auto","created_at":"2026-01-12 06:00:16","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":41734,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/615b85cedaa39350728aa5fd.png"},{"id":100008992,"identity":"86f4125c-ad49-4b6e-a028-ea0cf0ed6592","added_by":"auto","created_at":"2026-01-12 06:00:16","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":90529,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/94d7d6dc037447e4cd717df4.png"},{"id":100008969,"identity":"9ab33cbe-ccf7-4c84-9525-5be2cfa46fd8","added_by":"auto","created_at":"2026-01-12 06:00:14","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":25384,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/6c7b94fa8203a8faea94e86f.png"},{"id":100008978,"identity":"7dc0ddbb-e874-4387-85a4-22fa6432815f","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":150161,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/de7897ce8cf258a20bf00c5c.png"},{"id":100008981,"identity":"ce2f2494-0c1d-4b07-982b-90a0e5955782","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"xml","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":101473,"visible":true,"origin":"","legend":"","description":"","filename":"d4de8742548b463287955a45e5333caf1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/1891de0077e4fa9db5503bd3.xml"},{"id":100008973,"identity":"aba7891e-31e6-4459-9b71-2a959d821319","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"html","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":115072,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/b38d50d07b1a84276512cf41.html"},{"id":100008982,"identity":"3fb3bc2a-5948-4b84-baef-45397103c43c","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":184000,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverall workflow and characterization of fecal volatile organic compound (VOC) in colorectal cancer (CRC) diagnosis.\u003c/strong\u003e a: Schematic overview of the study design. Fecal samples from colorectal cancer (CRC) patients and healthy controls (HC) were collected, followed by VOC analysis using gas chromatography–ion mobility spectrometry (GC-IMS). The resulting data were subjected to a recursive feature elimination with cross-validation (RFECV) pipeline to select optimal features, which were then used to train and evaluate five machine learning models. b: Baseline demographic characteristics of the study cohort (n = 135 total), showing the numbers of males and females in the bar chart and the kernel density distribution of age in the lower plot. c: Distribution of the 85 identified fecal VOCs categorized by chemical class, based on Level 1 identification criteria\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/a2b95ea3e3709aefbcb3391b.png"},{"id":100361166,"identity":"ed6ca0c5-ea86-4f5d-92d0-0eead757d224","added_by":"auto","created_at":"2026-01-16 07:44:34","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":113469,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIdentification of key volatile organic compound (VOC) biomarkers using Recursive Feature Elimination with Cross-Validation (RFECV). \u003c/strong\u003ea–c: Cross-validation score curves illustrating the performance of the RFECV pipeline when coupled with different machine learning estimators: (a) Support Vector Machine (RFE-SVM), (b) Random Forest (RFE-RF), and (c) Least Absolute Shrinkage and Selection Operator (RFE-LASSO). The dashed red line in each panel indicates the optimal number of features selected by that specific estimator (e.g., 65 features for RFE-SVM). d: Venn diagram visualizing the overlap of the selected VOC features from all three RFECV methods. The central intersection reveals the 11 commonly selected key VOC biomarkers used for subsequent diagnostic model construction\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/05ace86ab1b54a96cb34d59f.png"},{"id":100360947,"identity":"79a79bcf-c866-4163-b3e6-20ae66dffe6a","added_by":"auto","created_at":"2026-01-16 07:44:13","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":358283,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMultidimensional evaluation of machine learning model performance for CRC detection. \u003c/strong\u003ea–b: Receiver operating characteristic (ROC) curves for the five models (XGBoost, CatBoost, AdaBoost, Random Forest, and LightGBM) on the (a) training set and (b) independent test set. The area under the ROC curve (AUROC) for each model is noted in the legend, demonstrating the model's discriminative ability. c–d: Precision-recall (PR) curves for the models on the (c) training set and (d) independent test set, evaluating performance on imbalanced data. E–F: Decision curve analysis (DCA) evaluating the clinical net benefit of the models across various threshold probabilities in the (e) training set and (f) independent test set, comparing the strategy against 'treat-all' and 'treat-none'. g: Brier Score comparison for each model on both training and test sets. Lower scores indicate better probability calibration (closer alignment between predicted probability and observed outcome\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/ebfd003e908ec19704dcda2d.png"},{"id":100360849,"identity":"ac3faa16-9caf-4a8e-8b4c-17869e57a0fa","added_by":"auto","created_at":"2026-01-16 07:44:02","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":91517,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eInterpretability analysis of the best-performing XGBoost model using SHapley Additive exPlanations (SHAP). \u003c/strong\u003ea: SHAP summary plot showing the global importance and directional impact of the 11 key features on the XGBoost model's output (colorectal cancer prediction). Features are ranked by their mean absolute SHAP value. The color scale indicates the original feature concentration (red = high, blue = low). b: SHAP dependence plot for the most influential metabolite, 3-Methylbutanoic acid-M, illustrating the relationship between its concentration (x-axis) and its SHAP value (y-axis). Higher concentrations of 3-Methylbutanoic acid-M generally correlate with lower SHAP values, suggesting this feature negatively influences the probability of a CRC diagnosis\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/1c690eed29da2010b67a992b.png"},{"id":100008988,"identity":"f55713c4-a48d-4a8b-854e-b88854304ffe","added_by":"auto","created_at":"2026-01-12 06:00:16","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":396449,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparative diagnostic efficacy and concentration differences of key volatile organic compounds (VOCs).\u003c/strong\u003e a: Receiver operating characteristic (ROC) curve for the clinically established tumor marker, carcinoembryonic antigen (CEA), showing an AUROC of 0.668 for discriminating colorectal cancer (CRC) from healthy controls (HC). b: ROC curves for the five most discriminative individual VOCs (Indole, 2-Phenylacetaldehyde, 3-Methylbutanoic acid-M, Propanoic acid-D, and 1-Butanol), demonstrating that the diagnostic performance of these single VOCs surpasses that of CEA. c: Violin plots depicting the concentration distribution of four specific VOCs (1-Butanol, 2-Phenylacetaldehyde, 3-Methylbutanoic acid-M, and Propanoic acid-D) between the CRC and HC groups. Statistical significance is indicated by asterisks (***\u003cem\u003eP\u003c/em\u003e\u0026lt; 0.001 and ****\u003cem\u003eP\u003c/em\u003e \u0026lt; 0.0001)\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/1f67e95826928cd511a97a3d.png"},{"id":109266354,"identity":"f81feab9-da9e-47db-98c8-66854b557308","added_by":"auto","created_at":"2026-05-14 12:40:45","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1266500,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/b824ad97-36dd-4cc2-9c55-e908c7528f31.pdf"},{"id":100008980,"identity":"29484cc6-a684-43af-a6d9-c33bae967b7f","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":178624,"visible":true,"origin":"","legend":"","description":"","filename":"11306mQACCminimumreportingstandardchecklist.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/12875049b31aeaf6b96b5169.pdf"},{"id":100008974,"identity":"5742d9a2-94d1-4023-90be-3a4f1bc2c30f","added_by":"auto","created_at":"2026-01-12 06:00:15","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":300411,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8505402/v1/2d88ef4e436c54abf79fd0cb.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Fecal volatile organic compound–based machine learning model for noninvasive detection of colorectal cancer","fulltext":[{"header":"Core Tip","content":"\u003cp\u003eThis study presents a novel, noninvasive diagnostic strategy for colorectal cancer (CRC) based on fecal volatile organic compound (VOC) profiling integrated with machine learning. Using GC\u0026ndash;IMS analysis and a robust feature selection pipeline, 11 key VOC biomarkers were identified. The XGBoost model achieved high diagnostic accuracy (AUROC = 0.8866) and superior clinical utility compared with carcinoembryonic antigen. SHAP interpretation revealed 3-methylbutanoic acid-M as the most influential biomarker. This VOC-based machine learning model offers a promising, accurate, and patient-friendly approach for early CRC detection and may improve compliance with large-scale screening programs.\u003c/p\u003e"},{"header":"INTRODUCTION","content":"\u003cp\u003eColorectal cancer (CRC) ranks among the most prevalent and lethal malignancies worldwide, representing a major global health challenge. According to the 2022 global cancer statistics report, CRC has the third-highest incidence and the second-highest mortality rate among all cancer(Bray et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The five-year survival rate for patients diagnosed at an early stage can reach 90%, but it declines sharply once the disease advances. Commonly used screening methods include fecal occult blood testing (FOBT), colonoscopy, carcinoembryonic antigen (CEA) measurement, and radiological imaging. Although colonoscopy remains the diagnostic gold standard for CRC(Rutter et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), its invasive nature, high cost, and requirement for bowel preparation restrict its utility for large-scale population screening(Quintero et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). FOBT, despite its convenience and noninvasive characteristics, suffers from limited sensitivity and specificity. Computed tomographic colonography (CTC) provides a noninvasive alternative but is less effective at detecting small or flat lesions and exposes patients to radiation while incurring substantial expense. Similarly, CEA serves as a general tumor marker for gastrointestinal malignancies, but its diagnostic role is auxiliary, with reported sensitivity and specificity of only 53% and 86%, respectively, precluding its use as a standalone diagnostic indicator(Acharya et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eVolatile organic compounds (VOCs), which are abundantly present in easily obtainable biological matrices such as exhaled breath, feces, and urine, have emerged as promising candidates for CRC screening because of their noninvasive nature, convenience, and low cost(Barbosa and Filho, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Li et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Markar et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). A meta-analysis of 32 studies reported that VOC-based testing achieved a sensitivity of 0.88 and a specificity of 0.85 in differentiating CRC patients from healthy controls (HC), with an area under the receiver operating characteristic curve (AUROC) of 0.93(Wang et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), providing strong evidence for its clinical potential. However, the high-dimensional and complex nature of VOC data presents significant analytical challenges. In this regard, machine learning (ML) offers an effective solution by identifying the most discriminative feature combinations from large datasets to develop robust diagnostic model(Handelman et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFurthermore, the value of ML is reflected in two major aspects: first, its ability to integrate multi-source heterogeneous data, thereby constructing more comprehensive and individualized diagnostic models; and second, its dynamic learning capacity, which enables real-time monitoring of disease progression and treatment response(Esteva et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). By capturing intricate, non-linear relationships within complex biological datasets, ML holds remarkable potential for early CRC detection, assessment of therapeutic efficacy, and prognosis prediction, capabilities that surpass those of traditional statistical approaches(Reel et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eHowever, the current field of VOC detection lacks standardized analytical protocols, resulting in a low concordance rate, less than 30%, of characteristic VOCs identified across different studies(Hanna et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), which directly undermines the comparability and reproducibility of findings. Moreover, VOC profiles are highly susceptible to external influences such as dietary habits, medication use, and environmental factors, suggesting that their diagnostic specificity still requires substantial improvement.\u003c/p\u003e \u003cp\u003eTo address the ongoing clinical challenges in early CRC diagnosis, this study aims to develop and validate an integrated diagnostic model that combines fecal VOC profiling with machine learning techniques. By optimizing both sample processing procedures and algorithmic workflows, we seek to identify key VOC biomarkers from high-dimensional data and construct a noninvasive diagnostic tool with superior accuracy compared to conventional tests. We anticipate that this approach will offer a novel, accessible, and high-performance solution to overcome the current limitations of CRC screening in sensitivity, specificity, and clinical applicability.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eOverview of participants and samples\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 135 participants were enrolled in this study at Huadong Hospital, Fudan University (Shanghai, China) between July 2024 and June 2025, including 78 patients with pathologically confirmed CRC and 57 HC. All subjects were recruited from the Yangtze River Delta region, a population characterized by relatively homogeneous dietary habits, thereby minimizing dietary confounding effects.\u003c/p\u003e\n\u003cp\u003eThe inclusion criteria were as follows: (1) age between 25 and 80 years; (2) histopathological confirmation of adenocarcinoma for CRC cases; and (3) willingness to participate and provision of written informed consent. HC participants were required to show no evidence of CRC or other organic gastrointestinal diseases, as verified by colonoscopy and imaging, and to have no history of malignancy or infectious disease.\u003c/p\u003e\n\u003cp\u003eExclusion criteria included: (1) pregnancy, lactation, or potential for pregnancy; (2) presence of congenital disorders; (3) history of psychiatric illness; (4) active peptic ulcer; (5) acute illness within two weeks prior to enrollment; (6) history of infectious diseases; (7) prior exposure to chemotherapy, radiotherapy, or combined oncologic treatments; (8) severe chronic obstructive pulmonary disease; (9) unstable diabetes mellitus; and (10) history of other malignancies.\u003c/p\u003e\n\u003cp\u003eAll participants provided written informed consent before sample collection. The study protocol was reviewed and approved by the Ethics Committee of Huadong Hospital Affiliated to Fudan University (Approval No. KY 2023K127).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eVOCs analysis methods and procedures\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eApproximately 0.5 g of each fecal sample was accurately weighed and transferred into a 20 mL headspace vial, followed by the addition of 1 mL of distilled water and 0.5 g of aspartic acid powder. The mixture was incubated at 60 \u0026deg;C for 5 min, after which 1 mL of headspace gas was extracted using a gastight syringe (Hamilton, Reno, NV, USA) and immediately injected into the analytical system, with nitrogen (99.999% purity) serving as the carrier gas.\u003c/p\u003e\n\u003cp\u003eVOCs were analyzed using a FlavourSpec\u0026reg; gas chromatography\u0026ndash;ion mobility spectrometry (GC\u0026ndash;IMS) system (G.A.S., Dortmund, Germany). Within this system, VOC components were initially separated through a GC column and subsequently introduced into the IMS detector for secondary separation and detection. Each VOC was identified in two dimensions, based on its GC retention index (RI) and IMS drift time (Dt), and quantified semi-quantitatively according to signal intensity.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eSelection of feature VOCs\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo identify key fecal VOC biomarkers associated with CRC, all data were normalized and subjected to feature selection using Recursive Feature Elimination with Cross-Validation (RFECV). This process was independently conducted using three core estimators: Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and a hyperparameter-optimized Support Vector Machine (SVM). Each algorithm was tuned and validated through 5-fold cross-validation to determine its optimal subset of discriminative features. The final core VOC panel was established by intersecting the selected features shared across all three algorithms, ensuring the robustness and reproducibility of biomarker selection.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eConstruction and evaluation of the predictive model\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eBased on the selected common VOC biomarkers, five machine learning models were constructed: eXtreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), Adaptive Boosting (AdaBoost), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM). The complete dataset was randomly divided into training and independent testing sets in a 7:3 ratio.\u003c/p\u003e\n\u003cp\u003eHyperparameter optimization was performed on the training set using the \u003cem\u003eRandomizedSearchCV\u003c/em\u003e function (Scikit-learn, Version 1.5.0; Python Software Foundation, Wilmington, DE, USA), sampling 20 random parameter combinations from a predefined grid. The optimization process was guided by the area under the receiver operating characteristic curve (AUROC), evaluated through 5-fold stratified cross-validation.\u003c/p\u003e\n\u003cp\u003eModel performance was comprehensively assessed using multiple indicators. On the training set, the mean and standard deviation of AUROC, area under the precision-recall curve (AUPRC), and Brier score were calculated via 5-fold stratified cross-validation. Decision curve analysis (DCA) was also conducted using the cross-validated predictions to estimate clinical utility. On the independent test set, the same metrics were computed, a confusion matrix was generated, and DCA was validated.\u003c/p\u003e\n\u003cp\u003eTo enhance interpretability, the best-performing model on the test set was further analyzed using SHapley Additive exPlanations (SHAP; Lundberg and Lee, University of Washington, Seattle, WA, USA), and the feature contributions were visualized through a beeswarm plot.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eStudy design and baseline characteristics of the study participants \u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eA total of 78 CRC patients and 57 HC were included in this study. Fecal samples were collected from all participants, and VOC data were obtained using GC\u0026ndash;IMS. Machine learning algorithms were subsequently applied to identify informative VOCs and construct a diagnostic model for CRC detection (Figure 1A).\u003c/p\u003e\n\u003cp\u003eBaseline demographic analysis demonstrated no statistically significant differences in age, sex, or body mass index (BMI) between the CRC and HC groups (\u003cem\u003eP\u003c/em\u003e \u0026gt; 0.05) (Table S1). As illustrated in Figure 1B, the proportions of male and female participants were comparable, indicating balanced gender distribution within the cohort. The age distribution profiles (violin bodies) of the two groups are highly overlapping, suggesting that the study population was well matched in terms of age structure.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eAnalysis of fecal volatile organic compounds\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eA total of 117 VOCs were detected in fecal samples. After excluding 32 compounds that could not be confidently identified, 85 VOCs were retained for subsequent analysis. Following the Level 1 identification criteria established by the Metabolomics Standards Initiative, these VOCs were classified into major chemical categories associated with CRC. The overall chemical composition consisted of alcohols (20%), fatty acids (17.65%), aldehydes (14.12%), ketones and esters (10.59% each), sulfur-containing compounds (5.88%), and other compounds (21.18%), including alkaloids, alkanes, furans, terpenes, and pyrazines (Figure 1C, Table S2). This classification reflects the chemical diversity of fecal VOCs and highlights multiple metabolite classes potentially implicated in CRC-related metabolic alterations.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eScreening results of feature VOCs\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe feature selection performance curves for RFE-LASSO, RFE-RF, and RFE-SVM are presented in Figure 2A\u0026ndash;C (Table S3). A Venn diagram (Figure 2D) illustrates the overlap among the VOC features identified by the three algorithms, highlighting 11 common VOCs consistently selected as key discriminatory biomarkers: (E)-2-pentenal, nonanal-D, 2-phenylacetaldehyde, indole, dimethyl trisulfide, 1-butanol, 4-methylphenol, pentanoic acid-M, propanoic acid-D, nonanal-M, and 3-methylbutanoic acid-M. These shared metabolites represent the most robust features contributing to CRC discrimination across multiple machine learning frameworks.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eVOCs-based XGBoost model for accurate diagnosis and interpretability analysis\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eTo comprehensively evaluate model performance, receiver operating characteristic (ROC) curves, precision\u0026ndash;recall (PR) curves, and decision curve analyses (DCA) were generated for both the training and test sets, along with comparative assessments of Brier scores.\u003c/p\u003e\n\u003cp\u003eDuring cross-validation on the training set (Figure 3A\u0026ndash;B), all models achieved ROC curves that were substantially higher than the random classification baseline. Among them, CatBoost demonstrated the highest AUROC, followed by LightGBM, XGBoost, RF, and AdaBoost in descending order. On the independent test set, all models maintained superior classification performance compared with random guessing. Notably, XGBoost exhibited excellent generalization ability, with a test AUROC nearly identical to its training result. At the optimal threshold, XGBoost achieved a sensitivity of 0.83 and a specificity of 0.78, reflecting a strong balance between detection and false-positive rates.\u003c/p\u003e\n\u003cp\u003eIn the PR analysis of the training set (Figure 3C), XGBoost achieved an AUPRC of 0.9121, ranking second only to CatBoost. However, in the test set (Figure 3D), its performance improved further, reaching an AUPRC of 0.9251, the highest among all evaluated models. DCA revealed that all models offered a substantially greater net clinical benefit than the \u0026ldquo;treat-all\u0026rdquo; and \u0026ldquo;treat-none\u0026rdquo; strategies across the 0\u0026ndash;0.8 probability threshold range (Figure 3E). On the test set, XGBoost demonstrated the highest net benefit within the high-threshold probability range (0.8\u0026ndash;1.0) (Figure 3F). Comparison of Brier scores (Figure 3G) showed minimal differences between the training and test sets across all models, with XGBoost achieving the lowest score, indicating superior calibration.\u003c/p\u003e\n\u003cp\u003eThe SHAP feature importance analysis for XGBoost (Figure 4A) identified 3-methylbutanoic acid-M as the most influential feature, exhibiting the broadest distribution and largest absolute SHAP values, followed by indole and 1-butanol. The SHAP dependence plot (Figure 4B) further demonstrated a positive relationship between the concentration of 3-methylbutanoic acid-M and the probability of CRC prediction, underscoring its critical contribution to model decision-making.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eAnalysis of diagnostic efficacy of biomarkers and expression differences of VOCs between CRC and HC\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eCompared with the clinically established tumor marker CEA (Table S4), which achieved an AUROC of 0.668 (Figure 5A), a value slightly higher than previously reported, its standalone diagnostic value for CRC remained limited. In contrast, several individual VOCs, including indole, 2-phenylacetaldehyde, 3-methylbutanoic acid-M, propanoic acid-D, and 1-butanol, demonstrated higher AUROC values than CEA (Figure 5B), indicating their superior diagnostic potential.\u003c/p\u003e\n\u003cp\u003eAs shown in Figure 5C, the concentrations of these metabolites differed significantly between groups. CRC patients exhibited markedly reduced levels of 1-butanol, 3-methylbutanoic acid-M, and propanoic acid-D compared with HC (\u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001), whereas the level of 2-phenylacetaldehyde was significantly elevated (\u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001). These findings highlight distinct metabolic alterations in fecal VOC profiles associated with CRC and underscore their potential as noninvasive biomarkers for early detection.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThis study aimed to establish a noninvasive diagnostic strategy for CRC using fecal metabolomic profiling integrated with machine learning. From 78 CRC and 57 HC samples, a total of 85 VOCs were identified through advanced analytical techniques. Employing an RFECV-based feature selection framework allowed the systematic extraction of robust biomarkers, effectively minimizing feature redundancy and selection bias often reported in earlier studies(Al-Tashi et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Zhang and Liu, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Among the models tested, XGBoost demonstrated the most consistent and superior performance across multiple evaluation metrics, particularly those relevant to clinical diagnostic accuracy. At the optimal decision threshold, the model achieved a sensitivity of 0.83 and a specificity of 0.78, meeting the clinical demands for both reliable detection and minimal misclassification(van Doorn et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The strong performance of XGBoost can be attributed to its gradient boosting structure, which iteratively refines residuals and dynamically adjusts feature weights, thereby capturing intricate non-linear interactions between VOCs and disease state. This finding is consistent with previous evidence highlighting the capacity of gradient boosting algorithms to handle high-dimensional biomarker data with exceptional precision(Guo et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Furthermore, SHAP analysis provided interpretability by elucidating the individual and combined effects of key VOCs on model predictions, thereby addressing the \u0026ldquo;black box\u0026rdquo; limitation of conventional machine learning models and offering mechanistic insight into the biological relevance of predictive features(Rudin, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWe identified 11 key VOCs, comprising three short-chain fatty acids (SCFAs), one sulfur-containing compound, four aldehydes, two alcohols, and indole. Comparative analysis against the widely recognized CRC screening biomarker CEA demonstrated that 1-butanol, 2-phenylacetaldehyde, 3-methylbutanoic acid-M, and propanoic acid-D exhibited significantly superior discriminatory power for CRC detection, underscoring their potential utility as dynamic indicators for disease monitoring.\u003c/p\u003e \u003cp\u003eThe reduction in SCFAs and their derivatives, including propanoic acid-D, pentanoic acid-M, and 3-methylbutanoic acid-M, was consistent with findings from previous studies(Ou et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Weir et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). Lower SCFA concentrations may compromise intestinal barrier integrity and immune homeostasis, thereby facilitating CRC development and progression(Silva et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Moreover, these metabolites serve as sensitive indicators of gut microbial metabolic activity. For example, diminished levels of propanoic acid-D may reflect a decreased abundance of SCFA-producing bacterial taxa or impaired carbohydrate fermentation capacity within the gut microbiome(Cong et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAldehydes such as (E)-2-pentenal, nonanal-D, and 2-phenylacetaldehyde are likely generated as by-products of lipid peroxidation and oxidative stress. During oxidative stress, reactive oxygen species attack polyunsaturated fatty acids within cell membranes, initiating a cascade of lipid peroxidation reactions that yield diverse VOCs, including aldehydes and alcohols, which can modulate cytotoxicity and apoptosis(Boots et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2012\u003c/span\u003e; West and Marnett, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2006\u003c/span\u003e). Among these, 2-phenylacetaldehyde, a crucial intermediate in phenylalanine metabolism, may participate in the biosynthesis of aromatic compounds and contribute to the regulation of cellular signaling pathways(Debnar-Daumler et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Tieman et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2006\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eSulfur-containing compounds, such as dimethyl trisulfide, are primarily produced by gut microbiota through the degradation of dietary sulfur-containing amino acids and are closely linked to intestinal inflammation and oxidative stress(Kumar et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe observed alterations in fecal VOCs were consistent with changes previously reported in breath VOC profiles from our earlier study(Liu et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), suggesting a potential metabolic link between fecal and breath volatile signatures. This concordance supports the hypothesis that systemic and intestinal metabolic processes contribute jointly to CRC-associated volatilomic changes. Furthermore, accumulating evidence indicates that gut microbiota may influence CRC pathogenesis through multiple mechanisms, including the production of bioactive metabolites, induction of chronic inflammation, and modulation of cancer cell energy metabolism(Yang et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). These insights provide a valuable framework for future mechanistic investigations into the microbiota\u0026ndash;metabolite\u0026ndash;cancer axis.\u003c/p\u003e \u003cp\u003eThis study, however, has several limitations. The cohort was derived from a single center with a modest sample size, which may introduce selection bias and limit the generalizability of the findings. All samples were collected at a single time point, precluding assessment of longitudinal changes in VOCs during disease progression. Due to the limited cohort size, subgroup analyses based on CRC stage or molecular subtype could not be performed.Despite these limitations, this study introduces a promising and practical framework for noninvasive CRC diagnosis by integrating VOC profiling with machine learning-based modeling.\u003c/p\u003e \u003cp\u003eTo address these limitations, we propose three strategies for subsequent research. First, we have established a collaboration with one clinical center in Southwest China and plan to collect at least 400 samples. By comparing model performance metrics such as AUC and sensitivity across data from different centers, we will systematically evaluate the impact of external factors, further refine the model, and establish a quantitative correction framework for confounding variables. Second, we will improve sample collection protocols by conducting comparative validation of sampling at different time points within the same day, aiming to achieve end-to-end standardization from sample collection to detection. Third, we will supplement the analysis with CRC subtype stratification(Altomare et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2013\u003c/span\u003e) to clarify the model\u0026rsquo;s applicability boundaries across populations with different clinical stages and pathological types, thereby supporting its precise clinical translation.\u003c/p\u003e"},{"header":"CONCLUSION","content":"\u003cp\u003eThis study developed and validated a noninvasive diagnostic approach for CRC by integrating fecal VOC profiling with advanced machine learning modeling. Using GC\u0026ndash;IMS and a robust feature selection strategy, 11 key VOC biomarkers were identified, several of which demonstrated superior diagnostic performance compared with CEA. The XGBoost model exhibited strong discriminative power, calibration, and clinical utility, achieving a balanced sensitivity and specificity suitable for early CRC detection. Moreover, SHAP analysis provided mechanistic interpretability, revealing that specific VOCs, such as 3-methylbutanoic acid-M and indole, play dominant roles in disease classification. These findings highlight the potential of VOC-based machine learning models as accessible, accurate, and patient-friendly tools for CRC screening. Future work should focus on large-scale, multi-center validation and the development of portable platforms to enable real-world clinical application.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll participants provided written informed consent. The study was approved by the Ethics Committee of Huadong Hospital Affiliated to Fudan University (Approval No. 2023K127) and was registered in the Chinese Clinical Trial Registry (ChiCTR2300073117).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interests\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u0026nbsp;\u003c/strong\u003eLi XW and Li QJ conceived the study. Li ZB developed the methodology and created the visualizations. Wang XH and Yu BR contributed to data collection, Li QJ and Li ZB drafted the original manuscript. Li QJ prepared figures 1-3. Li XW prepared figures 4-5 and interpreted results. Li XW and XH.W. provided supervision throughout the project. All authors reviewed and edited the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupported by\u003c/strong\u003e\u003cstrong\u003e：\u0026nbsp;\u003c/strong\u003eThis work was supported by the by the Key Specialized Disease Project (2022, ZDZB2221) and Key department (2022, ZDXK2213) of Huadong Hospital.\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAcharya, A., Markar, S.R., Matar, M., Ni, M. and Hanna, G.B. (2017) Use of Tumor Markers in Gastrointestinal Cancers: Surgeon Perceptions and Cost-Benefit Trade-Off Analysis. \u003cem\u003eAnn Surg Oncol\u003c/em\u003e \u003cstrong\u003e24,\u003c/strong\u003e 1165-1173.https://doi.org/10.1245/s10434-016-5717-y\u003c/li\u003e\n \u003cli\u003eAl-Tashi, Q., Saad, M.B., Muneer, A., Qureshi, R., Mirjalili, S., Sheshadri, A., Le, X., Vokes, N.I., Zhang, J. and Wu, J. (2023) Machine Learning Models for the Identification of Prognostic and Predictive Cancer Biomarkers: A Systematic Review. \u003cem\u003eInt J Mol Sci\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e.https://doi.org/10.3390/ijms24097781\u003c/li\u003e\n \u003cli\u003eAltomare, D.F., Di Lena, M., Porcelli, F., Trizio, L., Travaglio, E., Tutino, M., Dragonieri, S., Memeo, V. and de Gennaro, G. (2013) Exhaled volatile organic compounds identify patients with colorectal cancer. \u003cem\u003eBr J Surg\u003c/em\u003e \u003cstrong\u003e100,\u003c/strong\u003e 144-50.https://doi.org/10.1002/bjs.8942\u003c/li\u003e\n \u003cli\u003eBarbosa, J.M.G. and Filho, N.R.A. (2024) The human volatilome meets cancer diagnostics: past, present, and future of noninvasive applications. \u003cem\u003eMetabolomics\u003c/em\u003e \u003cstrong\u003e20,\u003c/strong\u003e 113.https://doi.org/10.1007/s11306-024-02180-5\u003c/li\u003e\n \u003cli\u003eBoots, A.W., van Berkel, J.J., Dallinga, J.W., Smolinska, A., Wouters, E.F. and van Schooten, F.J. (2012) The versatile use of exhaled volatile organic compounds in human health and disease. \u003cem\u003eJ Breath Res\u003c/em\u003e \u003cstrong\u003e6,\u003c/strong\u003e 027108.https://doi.org/10.1088/1752-7155/6/2/027108\u003c/li\u003e\n \u003cli\u003eBray, F., Laversanne, M., Sung, H., Ferlay, J., Siegel, R.L., Soerjomataram, I. and Jemal, A. (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. \u003cem\u003eCA Cancer J Clin\u003c/em\u003e \u003cstrong\u003e74,\u003c/strong\u003e 229-263.https://doi.org/10.3322/caac.21834\u003c/li\u003e\n \u003cli\u003eCong, J., Zhou, P. and Zhang, R. (2022) Intestinal Microbiota-Derived Short Chain Fatty Acids in Host Health and Disease. \u003cem\u003eNutrients\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e.https://doi.org/10.3390/nu14091977\u003c/li\u003e\n \u003cli\u003eDebnar-Daumler, C., Seubert, A., Schmitt, G. and Heider, J. (2014) Simultaneous involvement of a tungsten-containing aldehyde:ferredoxin oxidoreductase and a phenylacetaldehyde dehydrogenase in anaerobic phenylalanine metabolism. \u003cem\u003eJ Bacteriol\u003c/em\u003e \u003cstrong\u003e196,\u003c/strong\u003e 483-92.https://doi.org/10.1128/jb.00980-13\u003c/li\u003e\n \u003cli\u003eEsteva, A., Robicquet, A., Ramsundar, B., Kuleshov, V., DePristo, M., Chou, K., Cui, C., Corrado, G., Thrun, S. and Dean, J. (2019) A guide to deep learning in healthcare. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e25,\u003c/strong\u003e 24-29.https://doi.org/10.1038/s41591-018-0316-z\u003c/li\u003e\n \u003cli\u003eGuo, C., Liu, Z., Fan, H., Wang, H., Zhang, X., Zhao, S., Li, Y., Han, X., Wang, T., Chen, X. and Zhang, T. (2025) Machine-learning-based plasma metabolomic profiles for predicting long-term complications of cirrhosis. \u003cem\u003eHepatology\u003c/em\u003e \u003cstrong\u003e81,\u003c/strong\u003e 168-180.https://doi.org/10.1097/hep.0000000000000879\u003c/li\u003e\n \u003cli\u003eHandelman, G.S., Kok, H.K., Chandra, R.V., Razavi, A.H., Lee, M.J. and Asadi, H. (2018) eDoctor: machine learning and the future of medicine. \u003cem\u003eJ Intern Med\u003c/em\u003e \u003cstrong\u003e284,\u003c/strong\u003e 603-619.https://doi.org/10.1111/joim.12822\u003c/li\u003e\n \u003cli\u003eHanna, G.B., Boshier, P.R., Markar, S.R. and Romano, A. (2019) Accuracy and Methodologic Challenges of Volatile Organic Compound-Based Exhaled Breath Tests for Cancer Diagnosis: A Systematic Review and Meta-analysis. \u003cem\u003eJAMA Oncol\u003c/em\u003e \u003cstrong\u003e5,\u003c/strong\u003e e182815.https://doi.org/10.1001/jamaoncol.2018.2815\u003c/li\u003e\n \u003cli\u003eKumar, R., Sykes, D.J., Band, V.I., Schaller, M.L., Patel, R., Vitvitsky, V., Sajjakulnukit, P., Singhal, R., Wong, H.K.A., Hourigan, S.K., Ichinose, F., Lyssiotis, C.A., Shah, Y.M. and Banerjee, R. (2025) Gut sulfide metabolism modulates behavior and brain bioenergetics. \u003cem\u003eProc Natl Acad Sci U S A\u003c/em\u003e \u003cstrong\u003e122,\u003c/strong\u003e e2503677122.https://doi.org/10.1073/pnas.2503677122\u003c/li\u003e\n \u003cli\u003eLi, X., Shi, L., Long, Y., Wang, C., Qian, C., Li, W., Tian, Y. and Duan, Y. (2024) Volatile organic compounds in exhaled breath: a promising approach for accurate differentiation of lung adenocarcinoma and squamous cell carcinoma. \u003cem\u003eJ Breath Res\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e.https://doi.org/10.1088/1752-7163/ad6474\u003c/li\u003e\n \u003cli\u003eLiu, Y., Ji, Y., Chen, J., Zhang, Y., Li, X. and Li, X. (2024) Pioneering noninvasive colorectal cancer detection with an AI-enhanced breath volatilomics platform. \u003cem\u003eTheranostics\u003c/em\u003e \u003cstrong\u003e14,\u003c/strong\u003e 4240-4255.https://doi.org/10.7150/thno.94950\u003c/li\u003e\n \u003cli\u003eMarkar, S.R., Wiggins, T., Kumar, S. and Hanna, G.B. (2015) Exhaled breath analysis for the diagnosis and assessment of endoluminal gastrointestinal diseases. \u003cem\u003eJ Clin Gastroenterol\u003c/em\u003e \u003cstrong\u003e49,\u003c/strong\u003e 1-8.https://doi.org/10.1097/mcg.0000000000000247\u003c/li\u003e\n \u003cli\u003eOu, J., Carbonero, F., Zoetendal, E.G., DeLany, J.P., Wang, M., Newton, K., Gaskins, H.R. and O\u0026apos;Keefe, S.J. (2013) Diet, microbiota, and microbial metabolites in colon cancer risk in rural Africans and African Americans. \u003cem\u003eAm J Clin Nutr\u003c/em\u003e \u003cstrong\u003e98,\u003c/strong\u003e 111-20.https://doi.org/10.3945/ajcn.112.056689\u003c/li\u003e\n \u003cli\u003eQuintero, E., Castells, A., Bujanda, L., Cubiella, J., Salas, D., Lanas, \u0026Aacute;., Andreu, M., Carballo, F., Morillas, J.D., Hern\u0026aacute;ndez, C., Jover, R., Montalvo, I., Arenas, J., Laredo, E., Hern\u0026aacute;ndez, V., Iglesias, F., Cid, E., Zubizarreta, R., Sala, T., Ponce, M., Andr\u0026eacute;s, M., Teruel, G., Peris, A., Roncales, M.P., Polo-Tom\u0026aacute;s, M., Bessa, X., Ferrer-Armengou, O., Grau, J., Serradesanferm, A., Ono, A., Cruzado, J., P\u0026eacute;rez-Riquelme, F., Alonso-Abreu, I., de la Vega-Prieto, M., Reyes-Melian, J.M., Cacho, G., D\u0026iacute;az-Tasende, J., Herreros-de-Tejada, A., Poves, C., Santander, C. and Gonz\u0026aacute;lez-Navarro, A. (2012) Colonoscopy versus fecal immunochemical testing in colorectal-cancer screening. \u003cem\u003eN Engl J Med\u003c/em\u003e \u003cstrong\u003e366,\u003c/strong\u003e 697-706.https://doi.org/10.1056/NEJMoa1108895\u003c/li\u003e\n \u003cli\u003eReel, P.S., Reel, S., Pearson, E., Trucco, E. and Jefferson, E. (2021) Using machine learning approaches for multi-omics data analysis: A review. \u003cem\u003eBiotechnol Adv\u003c/em\u003e \u003cstrong\u003e49,\u003c/strong\u003e 107739.https://doi.org/10.1016/j.biotechadv.2021.107739\u003c/li\u003e\n \u003cli\u003eRudin, C. (2019) Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead. \u003cem\u003eNat Mach Intell\u003c/em\u003e \u003cstrong\u003e1,\u003c/strong\u003e 206-215.https://doi.org/10.1038/s42256-019-0048-x\u003c/li\u003e\n \u003cli\u003eRutter, M.D., Beintaris, I., Valori, R., Chiu, H.M., Corley, D.A., Cuatrecasas, M., Dekker, E., Forsberg, A., Gore-Booth, J., Haug, U., Kaminski, M.F., Matsuda, T., Meijer, G.A., Morris, E., Plumb, A.A., Rabeneck, L., Robertson, D.J., Schoen, R.E., Singh, H., Tinmouth, J., Young, G.P. and Sanduleanu, S. (2018) World Endoscopy Organization Consensus Statements on Post-Colonoscopy and Post-Imaging Colorectal Cancer. \u003cem\u003eGastroenterology\u003c/em\u003e \u003cstrong\u003e155,\u003c/strong\u003e 909-925.e3.https://doi.org/10.1053/j.gastro.2018.05.038\u003c/li\u003e\n \u003cli\u003eSilva, Y.P., Bernardi, A. and Frozza, R.L. (2020) The Role of Short-Chain Fatty Acids From Gut Microbiota in Gut-Brain Communication. \u003cem\u003eFront Endocrinol (Lausanne)\u003c/em\u003e \u003cstrong\u003e11,\u003c/strong\u003e 25.https://doi.org/10.3389/fendo.2020.00025\u003c/li\u003e\n \u003cli\u003eTieman, D., Taylor, M., Schauer, N., Fernie, A.R., Hanson, A.D. and Klee, H.J. (2006) Tomato aromatic amino acid decarboxylases participate in synthesis of the flavor volatiles 2-phenylethanol and 2-phenylacetaldehyde. \u003cem\u003eProc Natl Acad Sci U S A\u003c/em\u003e \u003cstrong\u003e103,\u003c/strong\u003e 8287-92.https://doi.org/10.1073/pnas.0602469103\u003c/li\u003e\n \u003cli\u003evan Doorn, W., Stassen, P.M., Borggreve, H.F., Schalkwijk, M.J., Stoffers, J., Bekers, O. and Meex, S.J.R. (2021) A comparison of machine learning models versus clinical evaluation for mortality prediction in patients with sepsis. \u003cem\u003ePLoS One\u003c/em\u003e \u003cstrong\u003e16,\u003c/strong\u003e e0245157.https://doi.org/10.1371/journal.pone.0245157\u003c/li\u003e\n \u003cli\u003eWang, Q., Fang, Y., Tan, S., Li, Z., Zheng, R., Ren, Y., Jiang, Y. and Huang, X. (2024) Diagnostic performance of volatile organic compounds analysis and electronic noses for detecting colorectal cancer: a systematic review and meta-analysis. \u003cem\u003eFront Oncol\u003c/em\u003e \u003cstrong\u003e14,\u003c/strong\u003e 1397259.https://doi.org/10.3389/fonc.2024.1397259\u003c/li\u003e\n \u003cli\u003eWeir, T.L., Manter, D.K., Sheflin, A.M., Barnett, B.A., Heuberger, A.L. and Ryan, E.P. (2013) Stool microbiome and metabolome differences between colorectal cancer patients and healthy adults. \u003cem\u003ePLoS One\u003c/em\u003e \u003cstrong\u003e8,\u003c/strong\u003e e70803.https://doi.org/10.1371/journal.pone.0070803\u003c/li\u003e\n \u003cli\u003eWest, J.D. and Marnett, L.J. (2006) Endogenous reactive intermediates as modulators of cell signaling and cell death. \u003cem\u003eChem Res Toxicol\u003c/em\u003e \u003cstrong\u003e19,\u003c/strong\u003e 173-94.https://doi.org/10.1021/tx050321u\u003c/li\u003e\n \u003cli\u003eYang, Y., Misra, B.B., Liang, L., Bi, D., Weng, W., Wu, W., Cai, S., Qin, H., Goel, A., Li, X. and Ma, Y. (2019) Integrated microbiome and metabolome analysis reveals a novel interplay between commensal bacteria and metabolites in colorectal cancer. \u003cem\u003eTheranostics\u003c/em\u003e \u003cstrong\u003e9,\u003c/strong\u003e 4101-4114.https://doi.org/10.7150/thno.35186\u003c/li\u003e\n \u003cli\u003eZhang, Z. and Liu, Z.P. (2021) Robust biomarker discovery for hepatocellular carcinoma from high-throughput data by multiple feature selection methods. \u003cem\u003eBMC Med Genomics\u003c/em\u003e \u003cstrong\u003e14,\u003c/strong\u003e 112.https://doi.org/10.1186/s12920-021-00957-4\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Colorectal cancer, Volatile organic compounds, Fecal biomarkers, Machine learning, Noninvasive detection, Gas chromatography–ion mobility spectrometry, Diagnostic model, SHAP analysis, XGBoost, Early screening","lastPublishedDoi":"10.21203/rs.3.rs-8505402/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8505402/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBACKGROUND\u003c/h2\u003e \u003cp\u003eColorectal cancer (CRC) remains a major global health concern, ranking among the top causes of cancer incidence and mortality. Current noninvasive screening tools such as fecal occult blood tests and serum carcinoembryonic antigen (CEA) assays suffer from limited sensitivity and specificity, while colonoscopy, the diagnostic gold standard, is invasive and costly. Volatile organic compounds (VOCs), metabolic end-products influenced by gut microbiota and tumor metabolism, offer a promising avenue for noninvasive CRC detection when coupled with advanced computational modeling.\u003c/p\u003e\u003ch2\u003eAIM\u003c/h2\u003e \u003cp\u003eTo develop and validate a fecal VOC-based machine learning model for noninvasive CRC detection.\u003c/p\u003e\u003ch2\u003eMETHODS\u003c/h2\u003e \u003cp\u003eFecal samples from 78 CRC patients and 57 healthy controls were analyzed using gas chromatography\u0026ndash;ion mobility spectrometry (GC\u0026ndash;IMS). Recursive feature elimination with cross-validation (RFECV) integrating LASSO, random forest, and support vector machine identified key VOCs. Five machine learning algorithms were constructed and optimized, and their diagnostic performance, calibration, and clinical utility were evaluated. SHapley Additive exPlanations (SHAP) analysis was applied to interpret model predictions.\u003c/p\u003e\u003ch2\u003eRESULTS\u003c/h2\u003e \u003cp\u003eAmong 85 identified VOCs, 11 were consistently selected as discriminative biomarkers, including 3-methylbutanoic acid-M, indole, and 1-butanol. The XGBoost model achieved the best performance with an area under the receiver operating characteristic curve (AUROC) of 0.8866, sensitivity of 0.83, and specificity of 0.78. SHAP analysis revealed 3-methylbutanoic acid-M as the most influential metabolite in model predictions. Several individual VOCs, such as 2-phenylacetaldehyde and propanoic acid-D, outperformed CEA in discriminating CRC from healthy controls. Decision curve analysis demonstrated superior clinical net benefit for the VOC-based model compared with traditional screening markers.\u003c/p\u003e\u003ch2\u003eCONCLUSION\u003c/h2\u003e \u003cp\u003eIntegration of fecal VOC profiling with a machine learning model provides a promising noninvasive strategy for accurate CRC detection, potentially improving early diagnosis and screening compliance.\u003c/p\u003e\u003ch2\u003eTrial Registration\u003c/h2\u003e \u003cp\u003eChinese Clinical Trial Registry (ChiCTR), ChiCTR2300073117. Registered on July 1, 2023 expected completion on June 30, 2025. Available at https//www.chictr.org.cn/bin/project/edit?pid=200842\u003c/p\u003e","manuscriptTitle":"Fecal volatile organic compound–based machine learning model for noninvasive detection of colorectal cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-12 06:00:08","doi":"10.21203/rs.3.rs-8505402/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4c3c78b1-3f30-4302-8218-50991dec5e59","owner":[],"postedDate":"January 12th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-14T12:29:18+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-14T12:40:38+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-12 06:00:08","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8505402","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8505402","identity":"rs-8505402","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}