Respiratory Disease Detection From Respiratory Sound And Medical History Using Deep Learning Techniques

Abstract

Respiratory diseases are infectious diseases that affect the respiratory organ, usually the lung of the patient. Oftentimes, investigation errors occur in respiratory disease diagnosis. To solve this, many researchers use different machine learning(ML) and Deep Learning(DL) techniques. But the previous works use respiratory sound as a single factor to predict respiratory disease. However, there is a significant overlap between the sound and symptoms of respiratory diseases. Here, this thesis aims in using patient medical history data along with respiratory sound to predict respiratory disease. For this purpose, we have collected patient symptom data from St. Paul Hospital and retrieved the respiratory sound data from ICBHI publicly available repository. Those datasets were used for pulmonary disease and respiratory sound classification in which, the target from respiratory sound is used as a feature to the pulmonary disease classification. To improve the respiratory sound classification model, feature Extraction MFCC and Mel-spectrogram along with augmentation techniques Time-shift, VTLP, and Spec-Aug have experimented with a CNN model. Towards pulmonary pathosis, the DNN model is used for classification. And the model has been evaluated with 5-fold cross-validation. Through experiments, respiratory sound classifier model gives an accuracy of 73.15% without Spec-Aug augmentation. From these experiments, we have observed that Mel spectrogram features give a good accuracy furthermore when we integrate it with the MFCC feature it gives us a better result. Also the augmentation techniques Time-shift and VTLP have provided us fine results while on the contrary, Spec-Aug augmentation technique gives modest accuracy, when in fact Spec-Aug has a much better speed of execution than other augmentation techniques. Whereas the pulmonary pathosis classifier gives 97% accuracy. Altogether the respiratory disease prediction performance has been evaluated with the help of a health expert, and the result obtained shows that our work correctly classifies 7 out of 10 subjects. Eventually, our work revealed that using patient symptoms along with respiratory sound enriches respiratory disease prediction. In addition, respiratory sound classification can benefit from Mel-spectrogram features; apart from that, the Spec Aug augmentation technique is unsuitable