Multi-feature representation for fatty liver disease detection with breath sample analysis

The human breath has various components to show an individual’s health status or represent a concrete disease, such as kidney disease or diabetes mellitus. Moreover, electronic nose (e-nose) is a widely used method to evaluate the breath sample of people with various chemical sensors to reflect their current health situation or predict various illnesses. Thus, e-nose breath analysis is a convenient, low-cost approach at present. So far, many works focus on various disease detection, such as diabetes, lung cancer, and kidney disease via breath sample analysis. However, few studies aim to investigate multi-feature fatty liver disease detection by evaluating breath samples via the e-nose. In this study, we propose a multi-feature representation method to extract multiple features for diagnosing fatty liver from healthy candidates via breath sample analysis. In particular, two external features, i.e., low-dimensional and latent as well as one internal feature, i.e., channel are extracted from the breath sample, which are further concatenated before being applied to various classifiers for diagnosis. Experimental results indicate that our proposed approach can obtain solid performances (Accuracy of 72.39% with SVM) in detecting fatty liver disease compared to only applying single feature methods.