Approach and application of extracting matching features from E-nose signals for AI tasks

Machine learning algorithms involving feature engineering as their key steps play important roles in AI tasks. Both feature selection and extraction have their pros and cons, recent studies have tried to meet the expectation of obtaining advantages of both methods, and this strive is still ongoing. Here, we proposed an idea of matching features analysis (MFA) based on wavelets to extract features in one-dimensional signals. To evaluate the performance of this method, "breath-prints" with 8364 features from 109 lung cancer patients and 109 healthy controls are sampled by our self-developed e-nose system. Features extracted by our proposed method and 8 other conventional techniques are fed to 3 classifiers. Performances of these 27 classification models are then compared. It turns out that, by using features extracted by our proposed method, classifiers show the best performance in recognizing lung cancer patients from controls. Classification accuracy of 93% and AUC of 0.96 is finally achieved when using MFA-MLP. With relatively better interpretation and approximation, this matching feature analysis technique may have great potentials in features engineering of biomedical signals for AI tasks.