Lung Sound Classification using Enhanced MFCC, Histogram, and Data Mining

• Main Author: Ingco, Wally Enrico
• Format: text
• Published: Archīum Ateneo 2020
• Subjects: n/a
• Online Access: https://archium.ateneo.edu/theses-dissertations/473
• Institution: Ateneo De Manila University

Chronic illnesses such as respiratory diseases are among the major health threats globally. Tuberculosis (TB) is a major public health concern worldwide and the world’s second most common cause of death from infectious disease after HIV/AIDs. Fortunately, with the advent of the Internet of Things (IoT) and the Artificial Intelligence (AI) concept, health condition monitoring had become easier and more accessible to the public. Mel Frequency Cepstral Coefficient, a well-known speech recognition feature provides a promising solution in analyzing and classifying lung sounds. This study aims to design and implement an enhanced MFCC model for lung sound classification using MATLAB. The model will help classify four different lung sounds, with data input taken and classified one at a time. The goal of which is to augment human intelligence and not to replace the existing lung sound classification methods. The pre-recorded lung sounds were characterized, and the researcher proposed four eMFCC models with three varying designs. The data collected from feature extraction and data mining were evaluated by the machine learning algorithms SVM and KNN. Measures like sensitivity, specificity, and accuracy were tested to determine which model was superior. Results showed that in terms of performance metrics, KNN performed better than SVM in classifying lung sounds. Tested in three designs where the pre-emphasis was removed, and the original 44.1kHz data resampled. Model 3 using KNN sampled at a frequency of 12000Hz has reached an average accuracy of 96.92% and a blind-data accuracy of 93.33%. A specificity of 97.94% and a sensitivity of 93.83%, achieving a performance that is comparable with existing studies on lung sound classification.