Chemical Nose-Based Non-Invasive Detection of Breast Cancer Using Exhaled Breath

Abstract

Breast cancer (BC) is the most commonly occurring cancer in women and one of the leading causes of cancer death in women worldwide. BC mortality is related to early tumor detection, highlighting the importance of early detection methods. This work aims to develop a robust, accurate and highly reliable, non-invasive, low-cost screening method for early detection of BC in routine screening using exhaled breath (EB) analysis. For this, exhaled breath samples were collected from 267 women: 131 breast cancer patients and 136 healthy women. After collection, the samples were measured using a commercially available electronic nose. The signals obtained for each sample were first processed and then went through a feature extraction step. An SVM model was then optimized with respect to the accuracy matrix using a validation set by applying a Monte Carlo cross-validation with 100 iterations, with each iteration containing 20% of the data. The validation set results were 80, 94, 88, and 95% for recall, precision, accuracy, and specificity, correspondingly. Once model optimization had concluded, 22 unknown samples were analyzed by the model, and an accuracy, precision, and specificity of 91% was achieved.

Keywords: breast cancer; data analysis; exhaled breath; machine learning.

Figure 1

Injection of a sample into the EN using a disposable syringe.

Figure 2

A typical sample prior to (left) and post (right)-normalization. Each line represents a different sensor. The results are presented after the extraction of the effective signal.

Figure 3

Boxchart of all standardized features selected for the model. Healthy subjects’ results are darker colored and can be seen to be more narrowly spread and negative than the sick subjects’ results. The sample legend is provided in the Supplementary Materials, X represents the mean of the dataset.

Figure 4

PCA plots for all features (left) and the selected, standardized features (right). Data from healthy subjects is in red, and from sick subjects in blue. Additional plots (PCA-3 and PCA-4) are presented in the Supplementary Materials.

Figure 5

Confusion matrix for the train data set. Labels 0 and 1 represent healthy and sick subjects, correspondingly.

Figure 6

Confusion matrix for the test data set. Labels 0 and 1 represent healthy and sick subjects, correspondingly.