Deep Learning and SMOTEENN-based Univariate Feature Selection Approaches for Diabetes Classification

Abstract

Modern advancement has been made in the data science domain for the past fifty years, this made that diagnosis through data-driven models are possible in this era. One of the main concerns in most data-driven classification models is feature selection, and hence this research proposes a SMOTEENN-based univariate feature test, which combines the strengths and minimise the downsides of both techniques. As of today, diabetes mellitus is one of the popular chronic diseases that have no cure up until this date, hence deep learning-based models will be constructed to perform this task, together with the proposed feature test mentioned earlier. As results, both DNN and CNN models achieved accuracies of 90% and above. However, the processes of univariate feature test are repetitive and time-consuming; overfitting and underfitting also occasionally happens during the training process; Whilst it is stated thusly, the method is beneficial to the final performance as it efficiently reduced computational and time cost for this task. In order to further verify this test, a larger and more complex dataset is recommended to fully utilise the potentials of deep learning models. The use of multivariate feature test should also be explored and compared with this test.