Predicting Diabetes using CNN for Various Activation Functions: A Comparative Study

Diabetes mellitus is a chronic disease that produces blood glucose abnormalities. This needs to be monitored at regular intervals. Many machine learning models made significant contribution in the prediction of diabetes. One such deep learning model is the convolution neural network. It is a multi-layered network architecture that comprises convolution, pooling, output layers, and activation functions. Activation functions are used to learn and approximate the relationships between various features in the model for final output. This research paper focus on comparing the performance of the CNN model using four non-linear activation functions; sigmoid, tanh, ReLU, ELU for blood glucose prediction. The model uses a simglucose UVA/Padova dataset of 30 patients for experimental analysis. Additionally, the model presents a three-class classification problem, where the patient is classified as euglycemic, hypoglycaemic, and hyperglycaemic. Statistical measures like recall, precision, F-score, and accuracy are used to assess the performance of CNNs employing various activation functions. The experimental results show that CNN gives the maximum accuracy of 97.3% when applied with the ELU activation function.