Prediction of cardiovascular risk by measuring carotid intima media thickness from an ultrasound image for type II diabetic mellitus subjects using machine learning and transfer learning techniques

• Published in: The Journal of supercomputing Vol. 77; no. 9; pp. 10289 - 10306
• Main Authors: Lakshmi Prabha, P., Jayanthy, A. K., Prem Kumar, C., Ramraj, Balaji
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.09.2021; Springer Nature B.V
• Subjects: Blood; Cardiovascular disease; Compilers; Computer Science; Demographic variables; Diabetes; Heart diseases; Interpreters; Machine learning; Mathematical analysis; Mobile and Intelligent Sensing on High Performance Computing; Multilayer perceptrons; Parameters; Processor Architectures; Programming Languages; Risk; Sensitivity; Support vector machines; Thickness; Ultrasonic imaging; Carotid intima media thickness; Framingham risk score; Transfer learning technique; Machine learning; Pearson correlation
• DOI: 10.1007/s11227-021-03676-w
• ISSN: 0920-8542

Cardiovascular disease (CVD) is a fatal disease that causes increased death in developing and developed nations. Among the various reasons, the increase in carotid intima media thickness (CIMT) is also a significant reason for CVD. It is expected to increase the death rate due to CVD up to 24.2 million by 2030. In previous studies, CIMT alone has been considered to identify the risk of CVD. In the proposed research, along with CIMT, the Framingham risk score (FRS) parameter was also calculated for both diabetic and normal subjects, which gives an accurate prediction of cardiovascular disease. CIMT was measured in 55 normal subjects and 55 diabetic subjects using a highly efficient ultrasound scanning device. Framingham risk score (FRS) was calculated for the 110 subjects based on the obtained demographic variables and biochemical parameters. The receiver operating characteristics (ROC) curve was plotted for CIMT with FRS which showed a sensitivity of 73% for CIMT. ROC curve plotted for FRS with fasting blood sugar and postprandial blood sugar showed a sensitivity of 80% and 81%, respectively. The performance was calculated based on different classification techniques. Results showed that support vector machine and multilayer perceptron classifier was classified with greater accuracy of 83.3% for 110 subjects. Further to improvise the analysis, the image data of the 110 subjects are augmented to 1809 image data and transfer learning techniques were applied using VGG16 and greater accuracy of 99% was achieved.