Diabetes Mellitus Prediction Using Transfer Learning

Abstract

Over 25% of the elderly population suffer from diabetes. Diabetes has no cure but an early diagnosis can assist in reducing its effects. Previously, Machine Learning has proven to be effective for diabetes prediction. However, in the literature, barely any methods used the high learning capacities of Deep Learning (DL) techniques for diabetes prediction. Hence, in this study, we have proposed methods for diabetes diagnosis using Deep Learning (DL). All the attributes in the Pima Indian Diabetes Dataset (PIDD) are crucial for diabetes diagnosis. Since medical data is sensitive, the requirement of a non-biased classifier is of utmost importance. Thus, the goal of this study is to create an intelligent model that can predict the presence of diabetes without using dimensionality reduction techniques. A 4-layered Neural Network (NN) model was used where the hidden layers consist of 64 neurons. Testing and evaluation demonstrated that the model achieves an accuracy of 93.33% on the PIDD. Alongside this, the data was also converted into an image dataset to apply transfer learning to the PIDD dataset. The obtained results are significantly better than the ones obtained via experiments from previous studies. The CNN models produce 100% accuracy scores on the PIDD. The study proves that CNNs can be successful when being used on small medical datasets. Based on the results, we can also conclude that the proposed system can effectively diagnose diabetes.