Machine Hearing a Cognitive Service for Aiding Clinical Diagnosis

Auscultation is being used for screening and monitoring respiratory diseases and is performed using a stethoscope. Auscultation the detection of abnormal respiratory sounds requires skilled medical professionals or clinical experts for diagnosis and an early diagnosis is always recommended for getting a higher probability of both curing and recovery. Respiratory disease being the most common with high morbidity the biggest challenge faced is the scarcity of clinical experts, non-availability of experts in rural and geographically challenged regions. Auscultation is an essential part of the physical examination, real-time and very informative, but based on the auditory perception of lung sounds. This requires the clinician considerable expertise and the perception variability may lead to misidentification of respiratory sounds. In this work, we are proposing an objective evaluation approach using deep learning techniques to address the limitations of the existing approach, a machine hearing technique to aid clinical decisions. In this work, breath sounds are used for analysis. The wheeze and crackles are the indicators of underlying ailments like namely Pneumonia, Bronchiectasis, Chronic Obstructive Pulmonary Disease (COPD), Upper Respiratory Tract Infection (URTI), Lower Respiratory Tract Infection (LRTI), Bronchiolitis, Asthma, and healthy. These sounds were analyzed for classifying the 8 categories of pulmonary diseases. CNN and RNN architectures were used for the classification of respiratory diseases. Features like Mel Frequency Cepstral Coefficients are extracted from the breath sounds. These coefficients were used as the feature for training the CNN and RNN architecture. Data Augmentation techniques like time stretching and shifting were applied to handle the imbalance in the data set. The CNN architecture gave a better accuracy of 0.89 and RNN with a slightly low of 0.833. The proposed approach proves to be a successful solution for the classification. The accuracy can be further improved with real-time data. In the future, this can be extended to develop a machine hearing as a decision support system for the clinical experts.