Detecting Colorectal Adenomas and Cancer Using Volatile Organic Compounds in Exhaled Breath: A Proof-of-Principle Study to Improve Screening

Abstract

Introduction: Early detection of colorectal cancer (CRC) by screening programs is crucial because survival rates worsen at advanced stages. However, the currently used screening method, the fecal immunochemical test (FIT), suffers from a high number of false-positives and is insensitive for detecting advanced adenomas (AAs), resulting in false-negatives for these premalignant lesions. Therefore, more accurate, noninvasive screening tools are needed. In this study, the utility of analyzing volatile organic compounds (VOCs) in exhaled breath in a FIT-positive population to detect the presence of colorectal neoplasia was studied.

Methods: In this multicenter prospective study, breath samples were collected from 382 FIT-positive patients with subsequent colonoscopy participating in the national Dutch bowel screening program (n = 84 negative controls, n = 130 non-AAs, n = 138 AAs, and n = 30 CRCs). Precolonoscopy exhaled VOCs were analyzed using thermal desorption-gas chromatography-mass spectrometry, and the data were preprocessed and analyzed using machine learning techniques.

Results: Using 10 discriminatory VOCs, AAs could be distinguished from negative controls with a sensitivity and specificity of 79% and 70%, respectively. Based on this biomarker profile, CRC and AA combined could be discriminated from controls with a sensitivity and specificity of 77% and 70%, respectively, and CRC alone could be discriminated from controls with a sensitivity and specificity of 80% and 70%, respectively. Moreover, the feasibility to discriminate non-AAs from controls and AAs was shown.

Discussion: VOCs in exhaled breath can detect the presence of AAs and CRC in a CRC screening population and may improve CRC screening in the future.

Figure 1.

Representation of the data analytics that were applied in this study to discriminate advanced adenomas from negative controls. In this study, ComBat was applied as a batch effect correction technique and Isolation Forests were used to select the representative subset for the independent test set. GC-MS, gas chromatography–mass spectrometry; RF, Random Forest.

Figure 2.

Hierarchical modeling approach to discriminate between negative controls, AA, and NAA using 3 sequential binary Random Forest models. First, AA was discriminated against the combination of controls and NAA. Subsequently, the latter were discriminated. To add an extra layer of sensitivity, Model 2 was added. AA, advanced adenoma; NAA, nonadvanced adenoma.

Figure 3.

Flowchart of included subjects. AA, advanced adenoma; BBPS, Boston Bowel Preparation Scale; CRC, colorectal cancer; FIT, fecal immunochemical test; IBD, inflammatory bowel disease; NAA, nonadvanced adenoma.

Figure 4.

(a) The obtained ROC curve when comparing colorectal cancer with negative controls using the Isolation Forest procedure. An AUC ROC, sensitivity, and specificity of 0.7313, 67.3%, and 70% were obtained, respectively. (b) The obtained ROC curve for the independent test set of the Random Forest model comparing advanced adenoma cases with control cases using 10 discriminatory volatile organic compounds. The sensitivity and specificity were found to be 70% and 90%, respectively. AUC ROC = 0.727, AUC precision-recall = 0.662. AUC, area under the curve; ROC, receiver operating characteristic.

Figure 5.

Principal Coordinate Analysis score plot on a proximity matrix obtained from the RF model and subsequently transformed using unsupervised RF built as described under the Methods section. The model was based on 10 selected volatile organic compounds. Every point belongs to a single breath fingerprint (red: advanced adenoma cases; blue: control cases). The separation is observed on Principle Component 1, explaining 70.5% of the variance. RF, Random Forest.

Figure 6.

Importance of the 10 selected and identified volatile organic compounds. Negative values indicate higher concentrations in control samples; positive values indicate higher relative concentrations in advanced adenoma cases.

Figure 7.

Principal Coordinate Analysis score plot on a proximity matrix obtained from the RF model built and subsequently transformed using unsupervised RF. The model was based on the midlevel fusion of Model 1 (advanced adenoma vs control) and the hierarchical subsequent models A and B. Every point belongs to a single breath chromatogram (blue: control cases; red: nonadvanced adenoma cases; yellow: advanced adenoma cases). The separation is observed on Principle Coordinates 1 and 2 together, explaining 76.75% of the variance. RF, Random Forest.