A novel hybrid machine learning framework for the prediction of diabetes with context-customized regularization and prediction procedures

Abstract

This paper proposes a customized hybrid model of artificial neural network (ANN) and genetic algorithms for an efficient diabetes disease prediction framework. Our customized hybrid model uses an improvised technique of detecting the more visible patterns of relations between the variables. Initially, the input medical dataset is preprocessed using a novel normalization technique that works consistently for all degrees of skewness of data. Then, our proposed decision-making algorithm will correctly identify the degree of importance of each variable in influencing the output, and thus priority will be given to the variables that are deemed most important. This is then followed by the implementation of a regularization method that is custom-made for the prediction of diabetes. Such a customized regularization method is considered asymmetrical because the positive numbers are more favored compared to negative numbers, and this was decided based on the characteristics of the dataset. The proposed technique deals with missing numbers as a separate kind of entity compared to numerical entries and can adapt itself to a given dataset. The proposed customized hybrid model and its accompanying decision-making algorithm were applied to the Pima Indian Diabetes dataset sourced from the UCI Machine Learning Repository with an 80% prediction accuracy.