Detection Performance of Colorectal Cancer Through Exhaled Breath by Electronic Nose: A Case-Control Study

Abstract

INTRODUCTION:

Although colorectal cancer (CRC) screening has been incorporated into organized programs in many countries, a universally accepted noninvasive and efficient screening method remains unavailable. This study aimed to assess the diagnostic potential of volatile organic compounds in exhaled breath through electronic nose (eNose) for noninvasive CRC detection.

METHODS:

The Cyranose320 sensor device was used to collect and analyze breath samples. Supervised machine learning was applied to evaluate the diagnostic performance of the eNose in CRC detection, using a randomly assigned training and validation set. Two-thirds of the breath samples were used to train models, which were then validated on the remaining patients (external validation). Three machine learning methods were applied for classification: random forest, extreme gradient boosting (XGBoost), and quadratic discriminant analysis.

RESULTS:

A total of 105 patients with CRC and 101 healthy controls were included. After adjusting for baseline covariates (age, sex, smoking, and comorbidities), machine learning models based on volatile organic compound profiles could differentiate patients with CRC from healthy controls, achieving areas under the receiver operating characteristic curve (AUC) of at least 0.72 in both the training and validation sets. The final CRC classification models yielded AUCs of 0.93 for random forest, 0.88 for XGBoost, and 0.89 for quadratic discriminant analysis. Furthermore, eNose classified CRC by stage, with an AUC exceeding 0.70 for early and advanced disease.

DISCUSSION:

Exhaled breath analysis using an eNose may serve as a promising noninvasive method for CRC detection. Further studies with larger populations are needed to confirm its clinical impact.

INTRODUCTION

Colorectal cancer (CRC) remains a major global health challenge, ranking as the second leading cause of cancer-related deaths worldwide (1). In China, the burden of CRC is particularly severe, with the 2020 China Cancer Statistics Report documenting 555,000 new cases and 286,000 deaths, making CRC the second most prevalent and fifth most lethal malignancy in the country (2).

Most CRC cases develop gradually from precursor lesions, such as adenomatous polyps or sessile serrated lesions. This prolonged developmental window presents a crucial opportunity for early detection and intervention, which are essential for improving patient outcomes and reducing CRC-related mortality (3). Despite the established benefits of early screening, CRC screening rates remain alarmingly low in many countries, including China, where only 14% of individuals older than 50 years of age undergo colonoscopy, the gold standard for CRC detection (4,5). Low screening adherence is influenced by multiple factors, including socioeconomic disparities, limited public awareness, and restricted access to screening programs (6,7). Therefore, there is an urgent need for accessible, noninvasive, and patient-friendly screening methods to enhance early CRC detection and alleviate the associated economic and public health burden.

Currently, the fecal immunochemical test (FIT) is the most widely used noninvasive CRC screening tool. However, its diagnostic performance is suboptimal, with sensitivity ranging from 69% to 86% for CRC and only 33%–47% for advanced adenomas (8,9). These limitations result in missed diagnoses and unnecessary colonoscopies, increasing the economic and logistical strain on healthcare systems. Given that the primary goal of CRC screening is to detect early-stage lesions and prevent cancer progression, there is a pressing need for more accurate, noninvasive, and cost-effective screening modalities. Addressing this gap represents a critical direction in CRC research and has the potential to significantly affect global cancer prevention strategies.

Volatile organic compounds (VOCs) have emerged as promising biomarkers for disease detection, reflecting metabolic alterations associated with various pathological conditions, including cancer, inflammation, and microbial dysbiosis (10). VOCs are present in multiple biological specimens, such as exhaled breath, saliva, urine, and feces, and their analysis offers several advantages over traditional diagnostic methods. Exhaled breath analysis, in particular, has garnered significant attention due to its noninvasive nature, reproducibility, and potential for real-time results. Previous studies have demonstrated the efficacy of VOC analysis in diagnosing a wide range of conditions, including gastrointestinal diseases (11), pulmonary disorders (12,13), endocrine abnormalities (14), infectious diseases (15–17), and various cancers (18–20). These findings underscore the potential of VOC-based diagnostics as a transformative approach to disease screening and monitoring.

Gas chromatography-mass spectrometry (GC-MS) is the most widely used and relatively accurate technique for VOC analysis. VOCs are present in multiple biological specimens, with exhaled breath and fecal samples being most clinically relevant for CRC screening due to their noninvasive collection and established diagnostic accuracy. Compared with conventional methods such as FIT, VOC analysis offers unique advantages including real-time results, potential for early detection of precancerous lesions, and ability to monitor treatment response.

Notably, GC-MS has been actively explored for polyp and CRC screening in research settings (21–23). Multiple studies have demonstrated its ability to differentiate VOC profiles between patients with CRC, individuals with colorectal polyps (including advanced adenomas), and healthy controls across biological specimens such as feces, urine, and breath (21,22,24). Despite its analytical precision, the clinical application of GC-MS is limited by its high cost, technical complexity, and requirement for specialized expertise, making it impractical for widespread use.

By contrast, the electronic nose (eNose) presents an innovative alternative tailored for clinical accessibility. Unlike GC-MS, which identifies individual VOCs, eNose uses an array of cross-reactive electrochemical sensors to detect collective patterns of VOCs, combined with machine learning algorithms to classify these patterns into diagnostic categories. This innovative design eliminates the need for sample pretreatment or technical expertise, enabling rapid analysis (often within minutes) at a fraction of the cost of GC-MS. For clinical readers, this translates to a tool that can be deployed in routine clinics without specialized laboratories, addressing key barriers such as accessibility and turnaround time that limit GC-MS's real-world use. Clinical validation studies have further supported eNose's potential: It has shown strong performance in distinguishing patients with CRC from healthy controls using both breath and fecal VOC profiles, and even demonstrated utility in identifying premalignant polyps (25). Notably, its portability and ease of use make it particularly suitable for primary care settings, potentially overcoming the accessibility barriers associated with conventional CRC screening methods.

Equipped with an array of electrochemical sensors and a pattern recognition system, the eNose rapidly detects and differentiates complex odor patterns at a lower cost. Its application in exhaled breath analysis has demonstrated excellent sensitivity, specificity, and diagnostic accuracy in various cancers, making it a compelling candidate for CRC screening (26).

Building on these advancements, this study aims to evaluate the potential of Cyranose320 in detecting and distinguishing patients with CRC from healthy controls.

METHODS AND MATERIALS

Study design

This prospective case-control study was conducted at the Affiliated Hospital of Chengdu University of Traditional Chinese Medicine (Chengdu, Sichuan, China) between April 2023 and March 2024. The study included 105 patients with CRC and 101 healthy volunteers as controls. All patients with CRC were newly diagnosed and had not received chemotherapy or radiotherapy before exhaled breath sample collection. Patients who had undergone surgical resection were included only if samples were collected before surgery or during preoperative evaluation (median 3 days before operation, range 1–7 days) to ensure captured VOCs reflected native tumor metabolism. The research protocol adhered to the ethical principles outlined in the Declaration of Helsinki, and written informed consent was obtained from all participants before their inclusion. Ethical approval was granted by the Medical Ethics Committee of the Affiliated Hospital of Chengdu University of Traditional Chinese Medicine (Approval Number: 2023KL-001). In addition, the study was registered with the Chinese Clinical Trial Registry (ChiCTR; Registration Number: ChiCTR2300067917).

Sample size

Based on previous studies and related literature (27,28), the required sample size was calculated using an established equation $\text{SE}=\sqrt{\frac{\mathrm{C}\hspace{0.17em}\left(100-\mathrm{C}\right)}{\mathrm{n}}}$. SE represents the standard error, C denotes the classification accuracy rate, and n indicates the estimated sample size. To ensure statistical validity, SE was set to not exceed 4%, with a target classification accuracy (C) of 80%. Using these parameters, the minimum sample size required for the training cohort was estimated at 100 patients.

Inclusion and exclusion criteria

CRC was diagnosed according to the Chinese Code for the Diagnosis and Treatment of Colorectal Cancer (2023 Edition) (2), and tumor-nodal-metastasis staging was determined based on the AJCC Cancer Staging Manual, 8th Edition (2017) issued by the American Joint Committee on Cancer. A validated questionnaire was administered to all participants to systematically collect data on comorbidities (e.g., hypertension, diabetes, and other chronic conditions), genetic history of CRC or other malignancies, and lifestyle/environmental risk factors such as smoking and alcohol consumption. The questionnaire had been pretested and validated in previous studies to ensure reliability and applicability.

Inclusion criteria for patients with CRC.

1. Pathologically confirmed CRC diagnosis.

2. Age 18 years or older, with no gender restrictions.

3. Voluntary participation with signed informed consent.

Exclusion criteria for patients with CRC.

1. Presence of hereditary nonpolyposis CRC, familial adenomatous polyposis, or inflammatory bowel disease.

2. Diagnosis of malignant tumors other than CRC.

3. Comorbidities such as respiratory infectious diseases, oral diseases, liver cirrhosis, chronic renal insufficiency, intestinal obstruction, esophageal diverticulum, diabetic nonketotic hyperosmolar syndrome, or diabetic ketoacidosis.

4. Antibiotic use within the past month.

5. Pregnancy.

6. Severe mental disorders (e.g., perceptual or thought disorders).

Inclusion criteria for healthy controls

1. No history of any type of cancer.

2. No diagnosis of acute diseases within the preceding 3 months and no recent medical treatment.

3. Age 18 years or older, with no gender restrictions.

Exclusion criteria for healthy controls.

1. Inability to provide a valid sample or adhere to the breath sample collection protocol.

2. Unwillingness to sign informed consent.

Control participants were recruited from individuals without a history of cancer who visited the same hospital during the study period. This approach ensured that both patients with CRC and controls were drawn from similar demographic and clinical environments, minimizing potential confounding factors.

Exhaled breath collection

To minimize the impact of lifestyle factors on exhaled breath samples, participants were required to brush their teeth after dinner and fast after 8 pm The night before sample collection. They were also instructed to refrain from smoking and consuming strong-smelling foods (e.g., garlic, leeks, or onions) or stimulant beverages (e.g., alcohol, coffee, or strong tea) for at least 12 hours before breath sampling.

Breath samples were collected between 6:30 am and 8:00 am the following morning. Participants rinsed their mouths 3 times with 100 mL of water and remained seated for at least 15 minutes in a clean sampling room. They were asked to relax, wear a nose clip, inhale deeply through the mouth 3 times, and then exhale slowly into a Teflon bag (E-switch, 1L, China) with an attached filter, ensuring no nasal ventilation. Once the bag was full, the valve was closed, and the investigator recorded the patient number on the sampling bag. The samples were stored at room temperature in a black plastic bag to protect them from light. They were transported under similar conditions and analyzed within 2 hours of collection.

Exhaled breath analysis

Exhaled breath samples were analyzed using the Cyranose 320 (Sensigent, CA), a portable eNose device equipped with 32 polymer-based sensors, each exhibiting distinct sensitivities to various VOCs (29). The Cyranose 320 is designed to detect and quantify a broad range of VOCs associated with cellular metabolic processes, including short-chain fatty acids (e.g., butyrate and propionate), aldehydes (e.g., hexanal and nonanal), ketones (e.g., acetone), and hydrocarbons (e.g., pentane). These VOCs are biologically plausible markers for CRC because their production is linked to dysregulated lipid peroxidation, gut microbial metabolism, and altered energy metabolism in cancer cells—hallmarks of colorectal carcinogenesis (30–32). For instance, elevated aldehydes may reflect increased oxidative stress in tumor microenvironments, while changes in short-chain fatty acids could indicate gut dysbiosis, a known risk factor for CRC progression (24,33).

Statistical analysis

Baseline characteristics.

Descriptive statistics were used to analyze baseline characteristics. Normally distributed continuous variables were presented as means with SDs, while nonnormally distributed variables were expressed as medians with interquartile ranges. Differences between the study and control groups were assessed using independent samples t-tests (for normally distributed variables) or Mann-Whitney U-test (for nonnormally distributed variables). Categorical variables were summarized as frequencies (percentages), and differences were evaluated using the χ² test or Fisher exact test. A 2-sided P value <0.05 was considered statistically significant.

VOC analysis.

Ambient air correction of sensor signals was performed, followed by peak normalization to the most stable sensor and interarray variance reduction (34). Sensor data were normalized using the fractional difference model: R/R₀= (Rmax-Ro)/Ro, where R represents the system response to the sample gas, R₀ is the baseline reading, and Rmax is the maximum response. The reference gas flow rate was set to ultrapure nitrogen.

Machine learning was applied to differentiate patients with CRC from healthy controls. Three classification models were developed: random forest (RF), extreme gradient boosting (XGBoost), and quadratic discriminant analysis (QDA). RF combines multiple decision trees, where each tree depends on an independently sampled random vector with the same distribution across all trees (35). XGBoost is a scalable machine learning algorithm that aggregates weak learners into a strong classifier (36). It constructs multiple decision trees sequentially, with each tree predicting the difference between the target value and the cumulative prediction from previous trees, thereby refining model accuracy. QDA is an extension of linear discriminant analysis that enables nonlinear data separation. Unlike linear discriminant analysis, which assumes equal covariance matrices across classes, QDA allows for class-specific covariance matrices, making it more flexible (37).

A 10-fold cross-validation approach was used to evaluate model performance. The data set was randomly divided into a training set (70%) for model development and a test set (30%) for validation. Model performance was assessed using sensitivity, specificity, positive predictive value, negative predictive value, and the area under the curve (AUC) with 95% confidence intervals (CIs).

All statistical analyses were performed using R-4.3.2 (R Foundation for Statistical Computing, Vienna, Austria). The tidyverse package was used for data preprocessing, and the tidymodels package was used for model development.

RESULTS

Baseline characteristics

A total of 227 subjects were screened, and 206 participants (105 patients with CRC and 101 healthy controls) were enrolled between April 2023 and March 2024. Twenty-one individuals were excluded from the final analysis: 5 due to a diagnosis of other tumor types and 16 due to failed breath tests (Figure 1). The demographic and clinical characteristics of the study population are summarized in Table 1.

Figure 1.

Literature screening flow chart. CRC, colorectal cancer.

Table 1.

Basic characteristics of the study subjects

Characteristics	Control	CRC
Subjects, N (%)	101	105
Age (y), median (IQR)	52 (41.0–59.0)	65 (57.0–70.0)
Male sex, n (%)	37 (36.6)	66 (62.9)
BMI, kg/m2, median (IQR)	23.0 (20.9–25.3)	22.7 (21.5–24.6)
Hypertension, n (%)	14 (13.9)	37 (35.2)
Diabetes, n (%)	4 (4.0)	20 (19.0)
Smoking status, n (%)
Current smoker	12 (11.9)	35 (33.3)
Ex-smoker	8 (7.9)	17 (10.5)
Never smoker	81 (80.2)	59 (56.2)
Cancer type
Adenocarcinoma		85 (81.0)
Other		20 (19.0)
Stage CRC
Stage Ⅰ		7 (6.7)
Stage Ⅱ		15 (14.3)
Stage Ⅲ		31 (29.5)
Stage Ⅳ		52 (49.5)
Tumor site
Colorectum		3 (2.9)
Right hemicolon		27 (25.7)
Rectum		56 (53.3)
Left colon		19 (18.1)
Metastases CRC
Yes		42 (40.0)
No		63 (60.0)

BMI, body mass index; CRC, colorectal cancer; IQR, interquartile range.

Patients with CRC were older (mean age: 63.3 ± 10.0 years), with a higher proportion of men, current or former smokers, and a greater prevalence of hypertension and diabetes compared with the control group (P < 0.05). Among patients with CRC, 79% were diagnosed at advanced stages (stage III or IV), while 21% were diagnosed at early stages (stage I or II). The majority of CRC cases were histologically classified as adenocarcinoma, consistent with its global predominance.

CRC vs controls

A total of 206 breath samples were collected, including 105 from patients with CRC and 101 from healthy controls. The data set was randomly partitioned into a 70% training set (CRC: n = 74, controls: n = 71) and a 30% validation set (CRC: n = 31, controls: n = 30) to identify CRC-specific VOC profiles. To address baseline imbalance, we adjusted for imbalanced baseline covariates (age, sex, smoking status, hypertension, and diabetes) by incorporating them as additional features in our models. The detailed diagnostic accuracy metrics, including AUC, sensitivity, specificity, positive predictive value, and negative predictive value, are summarized in Table 2. The receiver operating characteristic curves for the 3 models, illustrating their discriminatory power, are presented in Figure 2.

Table 2.

Diagnostic accuracy of colorectal cancer vs controls

Analysis	AUC (95% CI)	Accuracy (%)	Sens (%)	Spec (%)	PPV (%)	NPV (%)
RF
Training	0.91 (0.84–0.97)	85	77	94	92	81
Validation	0.80 (0.71–0.88)	77	77	76	74	79
Final	0.93 (0.90–0.97)	85	81	90	88	83
XG Boost
Training	0.89 (0.81–0.96)	84	81	87	86	82
Validation	0.72 (0.65–0.83)	74	73	75	71	76
Final	0.88 (0.83–0.93)	86	87	86	85	87
QDA
Training	0.88 (0.81–0.96)	84	77	90	89	80
Validation	0.78 (0.70–0.86)	75	79	71	70	80
Final	0.89 (0.85–0.94)	85	75	94	93	80

AUC, area under the curve; XGBoost, extreme gradient boosting; CI, confidence interval; CRC, colorectal cancer; NPV, negative predictive value; PPV, positive predictive value; QDA, quadratic discriminant analysis; RF, random forest; Sens, sensitivity; Spec, specificity.

Figure 2.

ROC curve plus scatter plot of the final disease-specific model for CRC. CRC, colorectal cancer; ROC, receiver operating characteristic.

In the training set, all 3 models (RF, QDA, and XGBoost) achieved overall accuracy exceeding 80% and maintained good performance in sensitivity and specificity both of which were above 75%. All 3 models exceeded an AUC of 0.85: RF (0.91), XGBoost (0.89), and QDA (0.88), with RF outperforming the others in this metric.

In the validation set, the RF and QDA models maintained robust performance. The RF model achieved 77% sensitivity, 76% specificity, and 77% accuracy, with an AUC of 0.80. The QDA model reached 79% sensitivity, 71% specificity, and 75% accuracy, with an AUC of 0.78. The XGBoost model showed slightly lower performance with 73% sensitivity, 75% specificity, 74% accuracy, and an AUC of 0.72.

For the final models, all 3 achieved an AUC above 0.85: RF (0.93), XGBoost (0.88), and QDA (0.89). The RF model demonstrated 81% sensitivity and 90% specificity, the XGBoost model achieved 87% sensitivity and 86% specificity, and the QDA model, while less sensitive (75%), exhibited high specificity (94%).

Second outcomes

The eNose effectively distinguished CRC stages and healthy controls. For secondary outcomes, we initially planned to compare 3 classification models: RF, XGBoost, and QDA. However, due to limited sample size, the QDA model failed to converge and was excluded from the final analysis. Only RF and XGBoost models were retained for analysis, with their performance in predicting secondary outcomes presented in Table 3.

Table 3.

Secondary outcomes

Analysis	Cases	Control	Method	AUC	Accuracy (%)	Sens (%)	Spec (%)	PPV (%)	NPV (%)
Early CRC vs control	22	101	RF	0.76	80.70	88.20	47.60	88.10	45.50
			XG Boost	0.71	77.40	88.20	27.30	84.90	33.00
Advanced CRC vs control	83	101	RF	0.70	65.60	72.50	57.10	67.00	62.70
			XG Boost	0.74	68.60	62.00	72.80	72.50	62.10
Early vs advanced CRC	22	83	RF	0.73	75.70	32.50	92.20	41.20	83.00
			XG Boost	0.64	68.70	41.70	75.80	31.00	82.9
Metastasis CRC vs control	42	101	RF	0.71	73.60	88.20	38.10	77.60	57.10
			XG Boost	0.74	72.20	84.30	42.90	78.20	52.90
Right hemicolon vs control	27	101	RF	0.67	73.80	86.30	28.60	81.30	33.30
			XG Boost	0.65	75.40	84.30	42.90	85.00	42.90
Left hemicolon vs control	19	101	RF	0.80	85.20	96.10	30.00	87.50	55.6
			XG Boost	0.74	83.60	96.00	20.00	86.00	42.9.00
Rectum vs control	56	101	RF	0.71	70.90	82.40	50.00	75.00	60.90
			XG Boost	0.70	65.80	72.50	53.60	74.00	51.70

AUC, area under the curve; XGBoost, extreme gradient boosting; CRC, colorectal cancer; NPV, negative predictive value; PPV, positive predictive value; RF, random forest.

In the comparison of early-stage CRC vs controls, the eNose achieved an AUC of 0.76, an accuracy of 80.70%, a sensitivity of 88.20%, and a specificity of 47.60% by using the RF model. For the XGBoost, the corresponding values were an AUC of 0.71, an accuracy of 77.40%, a sensitivity of 88.20%, and a lower specificity of 27.30%. These metrics indicate the eNose's capability to distinguish early-stage CRC from controls, with RF demonstrating relatively better overall performance in this comparison.

For advanced-stage CRC vs controls, both models achieved moderate diagnostic performance. The RF model yielded an AUC of 0.70, while the XGBoost model showed a higher AUC of 0.74, with both values exceeding 0.70, and the latter demonstrated enhanced performance in detecting metastatic CRC cases.

In addition, the eNose demonstrated site-specific diagnostic capability, with higher accuracy for left hemicolon CRC in the RF model (AUC = 0.80) compared with right hemicolon CRC in the same model (AUC = 0.67). These results underscore the potential of eNose technology in detecting CRC across different stages, metastatic statuses, and anatomical locations.

DISCUSSION

This study demonstrates that exhaled breath analysis using the Cyranose320 is a promising, noninvasive diagnostic tool for CRC detection. The findings indicate that eNose technology effectively distinguished patients with CRC from healthy controls, achieving AUC values exceeding 0.85 across the RF, XGBoost, and QDA models. This suggests the presence of distinct VOC signatures in the exhaled breath of patients with CRC, which can be reliably identified by this technology.

Notably, although our research contributes new insights by focusing on the Chinese population, it builds on a growing body of evidence supporting VOC-based diagnostics in CRC. As highlighted by Gower et al (38), who summarized the current state of knowledge regarding VOC-based testing for CRC, such techniques hold significant potential for improving cancer outcomes through enhanced detection and reduced healthcare burdens. Similarly, Chandrapalan et al conducted a systematic review on VOC analysis to enhance FIT testing in FIT-negative symptomatic populations, concluding that VOC could serve as a reliable test for ruling out CRC—further reinforcing the validity of VOC signatures in CRC detection (39). Early work by Bosch S et al (40), also laid groundwork in this field, contributing to the foundational understanding of VOCs' role in CRC diagnostics.

This study investigated the clinical utility and potential of eNose technology in a specific population context. Importantly, eNose has already been validated for CRC detection in other populations: Bosch et al, for instance, evaluated fecal VOCs using eNose in a Western cohort and found it effective for early detection of CRC and adenomas, as well as for determining optimal polyp surveillance timing-supported by their observation that VOC profiles normalized after polypectomy. Our findings, which confirm eNose's efficacy across diverse populations, align with the growing recognition of VOC-based diagnostics as a transformative approach in oncology, offering a noninvasive, rapid, and cost-effective alternative to traditional screening methods.

The diagnostic potential of VOC profiling has been widely documented, with applications in various malignancies, including lung, breast, prostate, liver, and thyroid cancers (41–46), as well as gastrointestinal disorders such as inflammatory bowel disease, ulcerative colitis, infectious diarrhea, and necrotizing enterocolitis (47–49). Current VOC analysis methods fall into 2 primary categories: (i) chemical analysis using GC-MS, which provides high precision and enables the identification of individual VOCs. (ii) Pattern recognition techniques, such as eNose, which detect complex VOC mixtures derived from clinical samples.

For instance, Amal et al (32) reported a sensitivity of 73%, specificity of 98%, and accuracy of 92% in detecting gastric cancer using GC-MS, while Altomare et al (30) achieved an AUC of 0.852 for CRC detection using the same methodology. Similarly, van Keulen et al (50) demonstrated the potential of exhaled VOCs as noninvasive biomarkers for CRC and advanced adenomas, with AUC values of 0.84 (sensitivity 95%, specificity 64%) for CRC and 0.73 (sensitivity 79%, specificity 59%) for advanced adenomas. These studies collectively underscore the diagnostic versatility of VOC analysis, while highlighting the unique advantages of eNose technology, particularly its ability to rapidly detect complex VOC patterns without requiring extensive sample preparation.

The generation of disease-specific VOCs is a complex process influenced by multiple factors. In CRC, VOC metabolites may result from uncontrolled cellular proliferation within the tumor microenvironment, systemic metabolic alterations associated with carcinogenesis, and dysregulated biochemical pathways (51). In addition, VOCs in feces, which reflect the composition and activity of the gut microbiota, have been shown to correlate closely with intestinal pathology. Bond et al (24) used GC-MS to analyze fecal VOCs in 137 patients, achieving an AUC of 0.82 for differentiating colorectal adenocarcinoma from nonneoplastic conditions, with 87.9% sensitivity and 84.6% specificity.

In addition to fecal VOCs, exhaled breath VOCs have been recognized as common biomarkers for diagnosing intestinal diseases, demonstrating high specificity and sensitivity (50,52). Furthermore, urinary VOC profiles also play a key role in the early diagnosis of intestinal diseases. A systematic review and meta-analysis showed that urinary VOCs effectively distinguished CRC from controls, with a sensitivity of 84% (95% CI: 73–91%) and a specificity of 70% (95% CI: 63–77%).

These findings suggest that a multimodal VOC analysis approach—integrating breath, fecal, and urinary biomarkers—could enhance diagnostic accuracy and provide a more comprehensive understanding of CRC pathophysiology.

A comprehensive comparison of VOC analysis methodologies is challenging because of inherent differences between chemical analysis techniques (which identify individual VOCs) and pattern recognition techniques (which detect collective VOC profiles). GC-MS, while highly precise, is often limited by its cost, complexity, and need for specialized expertise. By contrast, eNose technology offers a rapid, cost-effective alternative particularly suited for clinical applications. However, its inability to identify individual VOCs remains a significant limitation, necessitating complementary approaches for biomarker validation. Integrating eNose technology with advanced machine learning algorithms could enhance diagnostic performance by detecting subtle patterns in complex VOC mixtures. Future studies should combine the strengths of these methodologies, leveraging GC-MS for biomarker discovery and eNose for clinical implementation.

Our study stratified patients with CRC into early-stage (T1-T2) and advanced-stage (T3-T4) disease based on the tumor-nodal-metastasis classification system. The Cyranose320 demonstrated robust diagnostic performance when using the RF model, with AUC values of 0.76 for early-stage CRC, 0.70 for advanced-stage CRC, and 0.73 for distinguishing between early and advanced stages. In addition, the eNose demonstrated higher AUC values for left-sided lesions (0.80) and rectal lesion (0.71) compared with right-sided lesions (0.67), indicating that regional differences in CRC biology might affect VOC profiles. These findings highlight the potential of eNose technology not only for CRC detection but also for insights into disease progression and localization.

To eliminate external influences such as food intake, we required study participants to fast for 8 hours before exhaled breath sample collection. The sample bags used for breath collection underwent standardized washing and evacuation procedures to prevent environmental gas contamination. The results of this study provide additional support for breath VOC analysis as a viable CRC screening approach. The noninvasive nature, cost-effectiveness, and high patient compliance associated with breath VOC analysis underscore its potential as an alternative CRC screening method. Notably, eNose technology addresses several limitations of current screening methods, including the invasiveness of colonoscopy and the low sensitivity of FIT, offering a potential tool to enhance early detection and improve patient outcomes.

Limitations

Several limitations of this study must be acknowledged. First, the single-center design and relatively small sample size may limit the generalizability of findings and detection of subtle metabolic variations. Future studies should explore the integration of breath analysis, include larger, multicenter cohorts, and explore VOC profiles across different CRC subtypes and stages.

Second, the absence of a reference group with other gastrointestinal diseases precludes assessing specificity against non-CRC conditions. External validation studies are necessary to confirm the diagnostic accuracy of breath VOC analysis and differentiate CRC from other gastrointestinal diseases.

In addition, further research is needed to investigate the VOC profiles of different CRC subtypes and stages. Explore the integration of breath analysis with other diagnostic modalities to enhance screening accuracy and clinical utility.

The results of this study demonstrate that eNose technology can distinguish not only patients with CRC from healthy controls but also early-stage from advanced-stage CRC. These findings underscore the potential of eNose technology to enhance early detection, improve patient outcomes, and reduce the healthcare burden of CRC.

Future research should focus on expanding the sample size and validating results in diverse populations, refining analytical techniques to optimize diagnostic performance. Combining breath VOC analysis with other diagnostic methods to enhance clinical applicability.

CONFLICTS OF INTEREST

Guarantor of the article:

Specific author contributions: Q.W.: conceptualization (equal), methodology (equal), formal analysis (lead), writing-original draft, writing-review and editing (equal); S.T.: visualization (equal), formal analysis (equal), data curation (equal); R.Z.: visualization (equal), formal analysis (equal); Z.L.: supervision (equal), methodology (equal); Y.C.: writing-review and editing (equal); X.H.: conceptualization (equal), writing-review and editing (equal); Y.F.: supervision (equal), revision article.

Financial support: This study was funded by the Natural Science Foundation of Sichuan Province (No. 2022NSFSC0670) and (No. 2023NSFSC1809). The fund sponsor had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Potential competing interests: None to report.

Data availability: The data that support the findings of this study are available from the corresponding author on reasonable request.

Study Highlights.

WHAT IS KNOWN

• ✓ Although colorectal cancer (CRC) screening has been incorporated into organized programs in many countries, a universally accepted noninvasive and efficient screening method remains unavailable.

WHAT IS NEW HERE

• ✓ In this study, we aimed to evaluate the diagnostic potential of volatile organic compounds in exhaled breath analyzed by an electronic nose (eNose) as a noninvasive method for detecting CRC.

• ✓ We used the Cyranose320 sensor device to collect and analyze breath samples from participants and applied advanced machine learning algorithms-including random forest, extreme gradient boosting, and quadratic discriminant analysis to assess the diagnostic performance of the eNose.

• ✓ Our findings demonstrate that exhaled breath analysis using an eNose holds significant promise as a noninvasive diagnostic tool for CRC.

ABBREVIATIONS:

AUC

areas under the receiver operating characteristic curve

BMI

body mass index

CI

confidence intervals

CRC

colorectal cancer

eNose

electronic nose

FIT

fecal immunochemical test

GC-MS

gas chromatography-mass spectrometry

IQR

interquartile range

NPV

negative predictive value

PPV

positive predictive value

QDA

quadratic discriminant analysis

RF

random forest

Sens

sensitivity

Spec

specificity

TNM

tumor-nodal-metastasis

VOC

volatile organic compound

XGBoost

extreme gradient boosting