Ensemble Learning Model for Classification of Respiratory Anomalies

Machine learning methods for classifying respiratory anomalies have provided low classification accuracy. So high classification accuracy is required in order to utilize respiratory sound for classification of respiratory anomalies. This study proposes a convolutional neural network (CNN)-based ensemble model to improve accuracy in classifying respiratory anomaly level into normal, crackle, wheeze, and crackle + wheeze using respiratory sound data. We present three CNN models, each based on each of the three image forms: mel-frequency cepstral coefficient, waveform, or scalogram images. Each model demonstrates higher classification accuracies only for certain types of anomalies. The proposed ensemble model combines the three models to present a single model to classify all four types of respiratory anomalies with high accuracy; the competitive performance of the model is determined using the International Conference on Biomedical and Health Informatics (ICBHI) score. The classification accuracy and ICBHI score are compared with AI models in previous studies to validate the proposed model. The proposed model is expected to provide insights into early diagnosis of respiratory diseases through accurate classification of respiratory sounds, thereby facilitating early treatment of otherwise fatal diseases.