AusculNET: A Deep Learning framework for Adventitious Lung Sounds Classification.

Auscultation with a stethoscope is the most conventional means of examining and obtaining first insights into cardiovascular and pulmonary disorders. Although being noninvasive and inexpensive, requires access to medical professionals for their optimal usage. With the recent COVID-19 episode, the importance of having an in-house means for quick pulmonary assessment has grown. Digital stethoscope is a first step towards that but would be more potent if combined with on-device/online analysis of the sensed acoustic signals. Advancements in acoustic signal processing and deep learning models can be combined to propose new solutions to lung sound classification which could be integrated with pervasive acoustic sensors. We propose an automated lung sound analysis framework AusculNET, combining three convolution neural network (CNN) layers and one long-short term memory (LSTM) layer in conjunction with signal preprocessing and log-Mel spectrogram from each candidate breath cycle. AusculNET helps to decode the transient information in the lung sound waveform, classifying four classes of adventitious lung sounds - normal, crackle, wheeze, or both (crackle and wheeze). The explorations are carried out on the ICBHI 2017 Challenge Respiratory Sound Database, having 6898 breath cycles. We achieve an accuracy of 58% and an ICBHI score (accounting for the sensitivity and specificity due to unequal class distribution) of 55%, comparable with state-of-the-art. Furthermore, with an eye towards real-time implementation, our 8-bits quantized (both weights and activation) network achieves significant reduction in model parameters and improvement in overall inference times, incurring negligible accuracy loss.