Diabetes mellitus is considered to be a chronic disease that makes millions of lives miserable. Type I and Type II diabetes mellitus are caused depending upon the insulin level. Detection and diagnosis of diabetes mellitus has been performed effectively by various plasma glucose tests in the current medical systems including A1C test, fasting plasma glucose test (FPG) and oral glucose tolerance test (OGTT). Though the current diagnosis systems provide better results, the problems like imperfect concordance between the tests reduces the diagnosis accuracy. Another method of diagnosis is from the retinal images as diabetes causes retinal blindness, but these type of systems generally requires specialists to perform operation and also slightly expensive. Hence in this paper, the mobile based diabetes diagnosis system is proposed by obtaining the features from the real time inputs of glucometer with multiple measurement clinical data and retinal image features. The optimal features are selected using an optimization technique while the statistical measures are calculated for the time series data features. These features are fused together and trained using neural networks called hierarchical extreme learning machines (HELM) NN for the generation of dataset with minimum error samples. Then the association rule mining algorithm called FP-growth is employed to generate rules for determining the associations between the different sets of data. Thus the type I and type II diagnosis mellitus is diagnosed effectively with better accuracy using the proposed mobile based diabetes diagnosis system.