Evaluation of neural network compression methods on the respiratory sound dataset

Abstract

Recently, the use of smart medical solutions has experienced significant growth and the area of Internet of Medical Things (IoMT) has been established as an independent field. Artificial intelligence-based analysis of physiological signal data has resulted in promising results. This paper aims to assess neural network compression possibilities applied to a respiratory classification problem. The experiment is carried out on an Nvidia Jetson TX2 edge device and a personal computer. Respiratory sounds are classified into 8 classes of disease using two distinct deep learning networks. The trained models are compressed using half-precision and 8-bit integer quantization methods, and the inference results are compared and analyzed based on predictive powers, memory footprint, and inference time. Traditional, interpretable models (SVM, KNN) are also compared to the former deep models. The results are promising, the compression techniques manage to decrease memory usage 8 times while experiencing a negligible decrease in model accuracy.