Framework for machine learning artefact removal and empirical mode decomposition for capnogram based asthma detection

Abstract

Capnography has received considerable attention owing to its important applications in assessing asthma and other pulmonary diseases. Monitoring abnormal changes in the recorded carbon dioxide waveform (i.e., capnogram signal) allows for detecting respiratory malfunctioning and thereby averting potential asthma attacks. Detecting asthma based on the non-stationary capnogram signal remains an open research problem. In this thesis, an automatic computational framework is proposed to detect asthma. The presented framework includes two main stages. The first stage is responsible for discarding the distorted segments of the recorded capnogram signals. This task was performed in previous studies either manually by visual inspection, using threshold-based or template matching methods. In the current work, a machine learning-based approach is presented to automatically classify artefact-free and distorted capnogram segments. For this purpose, different time- and frequencydomain features are proposed. The time-domain features include energy, variance, skewness, kurtosis, Hjorth parameters and mean absolute deviation (MAD). The frequency-domain features include the area under the magnitude Fourier spectrum in addition to the number of relatively high spectral peaks for a particular frequency range. Different classifiers are trained and tested using the most relevant features: Hjorth activity and MAD. These classification models include Support Vector Machine (SVM), Random Forest (RF), Decision Tree (DT) andNaive Bayes (NB) classifiers. The results showed that the SVM classifier can provide classification accuracy, specificity, sensitivity and precision of 89%, 91%, 87% and 92.1%, respectively. In addition, a multiple classifiers voting approach is proposed for this classification task. Using this cooperative classification, the specificity is increased from 91% to 94%. The second stage accepts the clean capnogram segments from the first stage and carries out the classification of healthy and asthmatic capnograms. The proposed features are based on Empirical Mode Decomposition (EMD) which is suitable for analyzing the non-stationary capnogram signal in addition to the variance of the raw signal. Unlike the traditional features, the proposed features are extracted from the frequency-domain representation of the signal’s first Intrinsic Mode Function (IMF). The results showed that the NB classifier can provide classification accuracy, specificity, sensitivity and precision of 96.5%, 97%, 96% and 97.18%, respectively.