LADA Diabetes Classification by Combining Genetic and Conventional Risk Factors

Latent Autoimmune Diabetes in Adults (LADA) is the second most prevalent genre of diabetes after Type 2 but rarely recognised as a complex disease within its own right. Amid limited research benchmarks on LADA in the state-of-the-art, this thesis proposes a three-stage analysis for an accurate classification of LADA, with each stage informing benchmarks for the proceeding stage of analysis. Stage one models LADA using conventional variables alongside individual machine learning algorithms, where a classification accuracy of 85.51%, sensitivity of 84.09%, and specificity of 86.93%, alongside a precision of 86.93%, a recall of 84.53% and an F1 score of 85.71% are achieved. Stage two builds an ensemble of machine learning techniques identified in stage one along with building a novel Neural Ensemble Model (NEM) of three efficient deep learning techniques to model LADA classification and informs variable importance using conventional variables. Stage two generated a prediction accuracy of 92.00%, sensitivity of 91.77%, and specificity of 92.23% alongside a precision of 92.23%, recall at 91.79% and an F1 score of 92.02%. Stage three further improves the classification accuracy of LADA by introducing carefully chosen genetic variables below and above Genome Wide Association Studies (GWAS) along with the conventional variables fed as pseudo-random input sequence into the NEM model. Three different subsets including conventional variables alone, conventional and genetic variables below GWAS, and above GWAS are empirically investigated to determine the most reliable identification of case, control, precision, recall and variable importance. The proposed NEM model achieves a greater prediction accuracy when utilising a combination of conventional variables and eleven genetic variables below GWAS levels, with an accuracy of 96.06%, sensitivity of 97.45%, and specificity of 94.68% alongside a precision of 94.68%, recall of 97.38% and an F1 score of 96.01%, ultimately achieving a 97.45% correct identification of a positive LADA cases.