Fusion of Manual and Deep Learning Analyses for Automatic Lung Respiratory Sounds Identification in Youth
Keywords:Respiratory Sounds, Model Fusion, Deep Learning, Signal Processing, Paediatric
Lung sounds contain important clinical information which can be used for identifying respiratory and/or lung disorders. Manual identification of respiratory events is time-consuming and prone to subjective errors. While several automatic respiratory event classification techniques have been proposed previously, they are mostly focused on the identification of respiratory sounds in the adult population. Though, this is challenging in youth as lung is developing till the age of 20 years old which affects the parameters of respiratory sounds. In this research, our goal is to develop techniques for respiratory sound classification in youth using the SPRSound dataset, which includes recordings of individuals from 0-18 years old. The objectives include binary and multi-class classification of respiratory events (objective 1) and recordings (objective 2). For objective 1, we extracted purified respiratory features using a convolutional neural network (CNN) as well as frequency and time domain features, statistical features, and patient demographics, while for objective 2, a mixed model of long short-term memory (LSTM) network and a gradient boosting classifier with a novel voting scheme is developed. The features which were significantly associated with the different respiratory sounds were used to train machine-learning models for classification purposes. We evaluated the models’ performance based on sensitivity, specificity, an average of sensitivity and specificity scores, and the F1-score. The final performance score is defined as the average of the AS and F1- score. Our proposed framework reached 0.91±0.03 and 0.82±0.03 in binary and multiple event classification, respectively. Also, the developed model reached 0.74±0.02 and 0.55±0.03 in ternary and multi-class recording sound classification. Finally, the performance of the overall framework is calculated based on the grand challenge definition and our proposed pipeline reached 0.74.