Backpropagation Approach Supported by Image Compression Algorithm for the Classification of Chronic Condition Diseases.

Abstract

Diabetes is one of the main public health chronic conditions that are potentially reaching epidemic proportions globally. Worldwide, the occurrence of these types of diseases are increasing sharply at a worrying degree, with death of around 18 million people every year from cardiovascular disease, for which diabetes and hypertension are major predisposing factors. Two major concerns are that much of this increase in Diabetes is predicated to be happened in developing countries, with a growing incidence of Type 2 Diabetes (T2D) at a younger age including some obese children even before puberty. However, in developed countries most people with diabetes are above the age of retirement. As such, understanding the aetiology of T2D is vital. It has been thought that T2D is resulting from the convergence of genetics, environment, diet and lifestyle risk factors; however, genetic susceptibility has been established as a key component of risk. Genome-wide association studies (GWAS) is a study design and analytic tool specifically developed for investigating the genetic architecture of human disease. The ultimate aim of GWAS is to identify the genetic risk factors for common complex diseases such as T2D. Traditional parametric statistical approaches such as linear modelling framework (e.g. logistic regression) have limited power for modelling the complexity of genotype-phenotype relationship that is characterized by non-linear interactions. These nonlinear interactions are necessary in discovering the aetiology of complex diseases. More specifically, the linear modelling model has some limitations such as examining each single nucleotide polymorphisms independently for the association to the phenotype ignoring the epistatic (gene-gene interactions) and non-genetics factors. This paper presents a novel approch based on the use of backpropogation technique inspired by image compression algorithm. The proposed classifier is fine-tuned for binary classification to predict those who could suffer from the disease among those who do not. Simulation results indicated that the proposed technique showed an area under the curve, true positive rate, true negative rate values of 0.92, 0.9 and 0.8 respectively when using 2500 hidden neurons.