Modular Point-of-Care Breath Analyzer and Shape Taxonomy-Based Machine Learning for Gastric Cancer Detection

Abstract

Background: Gastric cancer is one of the deadliest malignant diseases, and the non-invasive screening and diagnostics options for it are limited. In this article, we present a multi-modular device for breath analysis coupled with a machine learning approach for the detection of cancer-specific breath from the shapes of sensor response curves (taxonomies of clusters).

Methods: We analyzed the breaths of 54 gastric cancer patients and 85 control group participants. The analysis was carried out using a breath analyzer with gold nanoparticle and metal oxide sensors. The response of the sensors was analyzed on the basis of the curve shapes and other features commonly used for comparison. These features were then used to train machine learning models using Naïve Bayes classifiers, Support Vector Machines and Random Forests.

Results: The accuracy of the trained models reached 77.8% (sensitivity: up to 66.54%; specificity: up to 92.39%). The use of the proposed shape-based features improved the accuracy in most cases, especially the overall accuracy and sensitivity.

Conclusions: The results show that this point-of-care breath analyzer and data analysis approach constitute a promising combination for the detection of gastric cancer-specific breath. The cluster taxonomy-based sensor reaction curve representation improved the results, and could be used in other similar applications.

Keywords: breath analysis; electronic nose; gastric cancer; machine learning; screening.

Figure 1

The point-of-care device used in the study: (a) the design of the device, with a disposable mouthpiece inserted at the front; (b) the main blocks of the system.

Figure 2

Diagram of the data analysis process.

Figure 3

An example of the data after the preprocessing: (a) common features describing the curve; (b) curves of a GNP sensor; (c) curves of an analogue MOX sensor; (d) curves of a digital MOX sensor.

Figure 3

An example of the data after the preprocessing: (a) common features describing the curve; (b) curves of a GNP sensor; (c) curves of an analogue MOX sensor; (d) curves of a digital MOX sensor.

Figure 4

Cluster taxonomy of the responses from one gold nanoparticle sensor.

Figure 5

The overall accuracy of Naïve Bayes classifiers: mean values and 95% confidence intervals.

Figure 6

Sensitivity of Naïve Bayes classifiers: mean values and 95% confidence intervals.

Figure 7

Specificity of Naïve Bayes classifiers: mean values and 95% confidence intervals.

Figure 8

The area under the ROC curve of Naïve Bayes classifiers: mean values and 95% confidence intervals.

Figure 9

An example of the characteristic shapes used in a Naïve Bayes model: a taxonomy for GNP sensor responses cut at six clusters (a), taxonomies of two other GNP sensors cut at 10 and four clusters (b,c), and one MOXD sensor at five clusters (d); the dashed lines shows individual measurements, and the solid bold lines show the cluster-characteristic shapes.

Figure 10

The overall accuracy of Random Forests: mean values and 95% confidence intervals.

Figure 11

The overall accuracy of SVMs: mean values and 95% confidence intervals.