Evaluation of Pre-Trained CNN Models for Cardiovascular Disease Classification: A Benchmark Study

Authors

Sami Alrabie, Department of the Information Technology, Faculty of Computing and Information Technology, Jeddah 21589, King Abdul Aziz University, Saudi ArabiaFollow
Ahmed Barnawi, Department of the Information Technology, Faculty of Computing and Information Technology, Jeddah 21589, King Abdul Aziz University, Saudi ArabiaFollow

Abstract

In this paper, we present an up-to-date benchmarking of the most commonly used pre-trained CNN models using a merged set of three available public datasets to have a large enough sample range. From the 18th century up to the present day, cardiovascular diseases, which are considered among the most significant health risks globally, have been diagnosed by the auscultation of heart sounds using a stethoscope. This method is elusive, and a highly experienced physician is required to master it. Artificial intelligence and, subsequently, machine learning are being applied to equip modern medicine with powerful tools to improve medical diagnoses. Image and audio pre-trained convolution neural network (CNN) models have been used for classifying normal and abnormal heartbeats using phonocardiogram signals. We objectively benchmark more than two dozen image-pre-trained CNN models in addition to two of the most popular audio-based pre-trained CNN models: VGGish and YAMnet, which have been developed specifically for audio classification. The experimental results have shown that audio-based models are among the best- performing models. In particular, the VGGish model had the highest average validation accuracy and average true positive rate of 87% and 85%, respectively.

Recommended Citation

Alrabie, Sami and Barnawi, Ahmed (2023) "Evaluation of Pre-Trained CNN Models for Cardiovascular Disease Classification: A Benchmark Study," Information Sciences Letters: Vol. 12 : Iss. 7 , PP -.
Available at: https://digitalcommons.aaru.edu.jo/isl/vol12/iss7/55