CNN+SVM+KNN: An Ensemble Framework for Predictive Analysis of Respiratory Sounds

Abstract

Vital information about a patient's lungs can be gathered through their respiratory sounds. The, the examination of lung sounds is a tedious process that needs a lot of time and results of this depends on the medical knowledge and diagnostic expertise of the doctors. This research introduces an ensemble framework to identify lung sounds for disease identification. The framework is an integration of convolutional neural network (CNN), support vector machine (SVM) and K-nearest neighbor (KNN). Initially, pulmonary acoustic features are extracted from recorded breathing sounds which are used for identification of respiratory disorders. Pulmonary sound signals are subjectively classified as normal, parenchymal, or obstruction of the airway’s pathology. CNN is used for its high specificity, dynamic result generation, high sensitivity, and model uniformity which predict the same values on a single input sample. SVM is trained on lung acoustic signals waveforms and spectrogram pictures without providing lesion-based criteria thus resulting with greater accuracy and specificity for detecting lung conditions in patients. KNN classifies sounds through an instance-based learning method that utilizes the feature space's closest training examples. The performance results of this proposed ensemble model is validated and compared with conventional classifiers such as Naïve Bayes (NB), Random Forest (RF), Gradient boosting (GB), Logistic regression (LR) and state-of-art works based on accuracy (%) metrics. The results  notify that the proposed model attains the highest accuracy of 98.76% so it  is enhancing patient results and the effectiveness of healthcare in the field of respiratory diseases.