Non-invasive Diagnosis of Diabetes Using Chaotic Features and Genetic Learning

Diagnosis of Diabetes Mellitus (DM) involves an invasive procedure. A pinch of blood sample is extracted by piercing a small needle into the body. This blood sample is fed to an electronic apparatus for detecting the blood glucose range. It is painful and repeated to keep track of the human body's temporal blood glucose level. Researchers are investigating alternative procedures to detect DM without injecting a needle into the human body. This work proposes a novel scheme for diagnosing DM, which is non-invasive and thus not painful. The proposed work takes the digital image of the human retina as input for the purpose. The input images are subjected to appropriate pre-processing beforehand, extracting meaningful features. During feature extraction, the focus is on identifying the chaotic geometric features formed due to the several non-uniform alignments of thin blood vessels inside the image. The chaotic geometry reveals intra-variability among the two classes under consideration (DM and Healthy). Feature vectors are generated to contain this intra-variability. Further classification is performed using a genetic learning method that involves a backpropagation neural network with modified learning weight updation through a Genetic Algorithm. Satisfactory results are obtained by implementing the scheme on a suitable dataset. The overall accuracy rate stands at 81.5%, ideal for emerging solution-oriented research work.