IJFANS International Journal of Food and Nutritional Sciences

ISSN PRINT 2319 1775 Online 2320-7876

A Novel Multiheaded Convolution Neural Network with Multilevel SVM Architecture for classification of Diabetes

Main Article Content

Sofia Goel, Sudhansh Sharma

Abstract

Classification of diabetes is an area of major concern in the field of medical science. Diabetes is not a single disease rather a group of health disorders that is hazardous to human health if not detected on time. Diabetes can be classified under three classes based on blood glucose level(BG): hyperglycemia(BG≥180mg/dL), hypoglycaemia (≤70 mg/dL) and euglycaemia (>70 and <180). BG level is detected through (CGM) techniques which generate time-series data. Various deep learning approaches like Convolution Neural Network have been employed for the classification of diabetes due to their multi-layered architecture and efficient feature extraction. However, the learning capacity of CNN model is limited due to the multidimensional times series nature of the data. Simply, adding more layers to the model makes it more complex and time-consuming and reduces the efficiency of classification. In this study, we have tried to overcome these limitations. This research proposes a hybrid model of Multiheaded CNN and multilevel SVM(MHCNN-3-SVM). In the proposed model, three multiple heads extract the features of each time series independently and the fully connected layer is replaced with multilevel SVM which is a powerful classifier. Also, the proposed model is tuned with different variants of head size and kernel size for performance optimization. UVA/PADOVA dataset of 30 patients (10 adults,10 adolescents,10 children) has been used in this study for multilevel classification. The results of the experiment depict that the proposed hybrid model obtains the highest accuracy of 99.49% and outperformed the state of the art approaches.

Article Details