Empirical Mode Decomposition and Grey Level Difference for Lung Sound Classification Sugondo Hadiyoso 1* , Achmad Rizal 2 1 School of Applied Science School, Telkom University, Bandung 40257, Indonesia 2 School of Electrical Engineering, Telkom University, Bandung 40257, Indonesia Corresponding Author Email: sugondo@telkomuniversity.ac.id https://doi.org/10.18280/ts.380118 ABSTRACT Received: 27 April 2020 Accepted: 10 November 2020 Lung sound is one of the parameters of respiratory health. This sound has a specific character if there is a disease in the lungs. In some cases, it is difficult to distinguish one type of lung sound to another. It takes the expertise, experience and sensitivity of clinicians to avoid misdiagnosis. Therefore, many studies have proposed a feature extraction method combined with automatic classification method for the detection of lung disease through lung sound analysis. Since the complex nature of biological signals which are produced by complex processes, the multiscale method is an interesting feature extraction method to be developed. This study proposes an empirical mode decomposition (EMD) and a modified gray level difference (GLD) for a lung sound classification. The EMD was used to decompose the signal, and then GLD was measured on each decomposed signal as a feature set. There are five classes of lung sounds which were simulated in this study, including normal, wheeze, crackle, pleural rub, and stridor. Performance evaluation was carried out using a multilayer perceptron (MLP) and 3-fold cross-validation. This proposed method yielded the highest accuracy of 96.97%. This study outperformed several previous studies which were simulated on the same dataset. It is hoped that in the future, the proposed methods can be tested on larger datasets to determine the robustness of the methods. Keywords: lung sound, EMD, GLD, MLP 1. INTRODUCTION Lung sounds can provide information on lung health [1]. Lung sounds can be heard using a stethoscope, which is called the auscultation method [2]. If there are physiological changes in the lungs due to a disease, it can cause changes in the lung sound pattern [3]. This becomes one of the reference criteria for doctors to diagnose the patients' lung disease. However, this technique tends to be very subjective because it depends on the experience and expertise of the doctor [1]. Digital signal processing methods become an alternative to deal with that problem. Many studies have proposed algorithms for automatic classification of lung sounds based on digital signal processing. Various feature extraction methods combined with classifier algorithms have been reported. Some of them use time-domain analysis, frequency domain analysis, time-frequency domain analysis, and signal complexity analysis. Analyzes in the time domain, for example using the Hjorth descriptor method are reported in [4-6] or empirical mode decomposition method, as reported in the study by Chen et al. [7]. Other studies that use analysis in the frequency domain such as Fast Fourier Transform (FFT) [8, 9], or using the Mel Frequency Cepstral Coefficients (MFCC) method are reported in the studies [10, 11] Lung sound classification using analysis in the time- frequency domain, as reported in the studies [12, 13]. However, time domain analysis is preferred in the case of short data segments and it is capable of measuring the power over the broad band into a single index [14]. Time domain technique is thought to be more suitable in the case of abnormal lung sounds, because it has a short period [15, 16]. Since naturally the lung sound signals have non-stationary properties, a complexity analysis in the time domain is used to characterize this signal by some researchers. The most common complexity approach is the entropy measurement. Previous studies by Rizal et al. proposed seven combinations of entropy measurements for a lung sound classification [17]. This study generated an accuracy of 94.95% for the classification of five classes of lung sounds. Another entropy- based research on the classification of wheeze and non-wheeze within lung sound signals was proposed by Aydore et al. [18]. The Renyi entropy method was used for feature extraction, and the accuracy achieved was 93.5% for the two data classes. Another study used a complexity parameter based on the Hjorth method for the feature extraction process. Hjorth parameters were measured on a single scale and multiscale scheme. The proposed methods generated an accuracy of 83.95% and 95.06%, respectively [4, 5]. From the literature review on previous studies where the complexity approach has been performed, there is still a gap to improve the accuracy. Therefore, in this study, we proposed new protocol based on the empirical mode decomposition and a modified grey level difference (GLD) for lung sound classification. In this method, GLD parameters were measured at each level of decomposition. The calculated GLD parameters included second-moment gradient (GSM), contrast gradient (GC), mean gradient (GM), inverse difference moment (IDM), and gradient entropy (GE). These features then become the input of a multilayer perceptron for lung sounds classification. It is expected that this GLD can produce higher accuracy compared to entropy and Hjorth methods. As a reminder, this paper is organized as follows: section 2 Traitement du Signal Vol. 38, No. 1, February, 2021, pp. 175-179 Journal homepage: http://iieta.org/journals/ts 175