Machine Learning Approaches for Diabetes Mellitus Prediction and Management

The disease of diabetes mellitus (DM) occurs due to elevated blood glucose levels in the bloodstream. Early prediction of DM progression can bring substantial health benefits for the patients and facilitate proactive care. In recent years, advanced prediction of type 2 diabetes (T2DM) was not given enough focuses in the literature despite its significant implication on patients’ health. Moreover, the extraction of biomarkers which are significantly correlated with the future progression of T2DM still remains unexplored. Furthermore, there is no previous work for advanced prediction of hemoglobin A1c (HbA1c), which is extensively used for DM management and forecasting complications. This dissertation aims for early onset detection of T2DM by developing a prediction model that identifies individuals at risk of developing diabetes in the future. This dissertation also aims for the long term prediction of HbA1c levels using a multi-stage multi-class (MSMC) data analysis approach. The proposed framework comprises novel methods for missing data estimations, feature extraction, feature selection, and implementation of machine learning (ML) and deep learning (DL) model for early onset detection of T2DM and HbA1c prediction. The developed T2DM prediction framework is evaluated using 1368 patients’ oral glucose tolerance test (OGTT) data sourced from the San Antonio Heart Study. Furthermore, to evaluate the developed HbA1c prediction framework, a total of 200 patients’ 3000 days continuous glucose monitoring (CGM) data collected from Sidra Medicine, Doha, Qatar, have been analyzed. Our proposed fractional derivative and Haar wavelet transformation feature extraction and ensembling of naïve Bayes, support vector machine, and random forest classifiers achieve an accuracy of 95.94% for T2DM prediction. Furthermore, the conversion of CGM data into binary and histogram images and a few-shot learning distance (FSLD) feature extraction approach shows an accuracy of 92.30% in advanced prediction of HbA1c levels. Our developed framework on early onset detection of T2DM outperforms state of the art in accuracy and other evaluation metrics. For the first time in the literature, advanced HbA1c prediction is attempted. The significance of this research work is crucial because it allows subjects to be given a fair warning of whether they are susceptible to developing T2DM in the future. This early warning can help prevent the disorder by taking appropriate measures and, at minimum, reducing the severity of the disease and prolong its onset.