Accurate respiratory sound classification model based on piccolo pattern. (January 2022)

• Record Type: Journal Article
• Authors: Tasar, Beyda; Yaman, Orhan; Tuncer, Turker
• Is Part Of: Applied acoustics. Volume 188(2022)
• Publication Date: 2022-01
• Subjects: Pulmonary diseases -- Lung sound analysis -- Piccolo pattern -- NCA feature selection -- TQWT; Acoustical engineering -- Periodicals; Periodicals; 620.2
• DOI: 10.1016/j.apacoust.2021.108589
• Languages: English
• ISSNs: 0003-682X

Abstract

Highlights: Nonlinear histogram-based local feature generation model is presented. Multilevel fused feature generation based sound signal classification model is proposed. This model reached 99.45%, 99.31%, and 99.19% accuracies on three cases. Piccolo pattern based model outperformed. Abstract: Auscultation, or listening to body sounds with a stethoscope, is the most basic and instructive part of the physical examination. Although it is a quick and effective way of diagnosing respiratory tract disorders, the precision of the diagnosis demands a considerable deal of clinical knowledge. As a result, we've created a computerized diagnostic system that will detect breathing sounds automatically, assisting physicians and trainees in the specialization process. Automated respiratory sound classification is a complex issue for advanced biomedical signal processing. Therefore, variable models and methods have been introduced to overcome this problem. This research aims to introduce a high accurate sound classification model using a nonlinear histogram-based generator. To reach this aim, piccolo pattern (it uses S-box of the piccolo cipher as a pattern), statistical moments, tunable q-factor wavelet transform (TQWT), iterative neighborhood component analysis (INCA), and conventional classifiers are used together. This model uses TQWT to create levels. Piccolo pattern is employed to generate textural features (it is the main feature generator of this model), and statistics are deployed to extract statistical features. INCA chooses the relevant features. Decision tree (DT), support vector machine (SVM), and k nearest neighbors (KNN) classifiers are applied to the selected feature vectors to calculate results. Three cases are defined using ICHBI 2017 dataset to calculate results comprehensively. The defined cases contain seven, three, and eight categories, respectively. Furthermore, five performance calculation metrics are employed to evaluate the presented piccolo-pattern-based model comprehensively. The introduced piccolo pattern-based model reaches 99.45%, 99.31%, and 99.19% accuracies for Case 1, Case 2, and Case 3 by employing KNN. These accuracies denote the success of the piccolo pattern-based model. Furthermore, this model attains higher accuracies than the previously presented deep learning-based methods for respiratory sound classification.