| Diabetes has become a major global public health problem. Diabetic patients have very high morbidity rates of cardiovascular complications. Those complications affect their quality of life and are the major causes of blindness, limb amputation and mortality. This thesis attempts to construct effective models in predicting chronic type-2-diabetes-related complications in aims for better control and treatment of those diseases.Data Mining refers to extracting meaningful and potentially useful information from a large amount of incomplete and noisy data. Data mining has been popular in many applications. In clinical diagnosis, data mining approaches can be adopted to construct prediction models from clinical and laboratory data. These models can improve diagnostic accuracy and provide evidences and guidance for decision making.In this thesis, three methods are investigated for constructing prediction models of T2D complications. The first method pre-processes the collected data, and then builds a model through Learning Vector Quantitative (LVQ) neural network, which is an adaptive clustering approach using the labeled data as the training set. The second method uses LOGISTIC Regression to construct a statistical model. The third method uses LVQ neural network to construct a model based on the variables that are selected from LOGISTIC Regression. It is of interest to determine how the clinical and laboratory data are related to the T2D complications and to establish the prediction models for the chronic complications of T2D.The three methods above are used to predict five chronic T2d complications: diabetic cardiovascular disease (DC), diabetic nephropathy (DN), diabeticretinopathy (DR), arterial embolism in lower limb and diabetic nerve complication. The results show that LVQ method has classification accuracy rates of 70.59%, 78.43%, 74.51%, 90.20% and 68.63% respectively for the above five complications. The corresponding accuracy rates of logistic regression model are 76.47%, 78.43%, 62.75%, 86.27% and 60.78%. The classification accuracy rates of the third method are 70.59%, 78.43%, 70.59%, 84.31% and 64.71%.As a conclusion, all of the three methods provide fairly good prediction of the complications DC, DN, DR and arterial embolism, although they are unable to predict well the diabetic nerve complication. The three prediction methods have no statistically significant difference in prediction accuracy. The prediction models developed in this thesis are valuable for improving the treatment of the type 2 diabetes patients and for preventing the possible complications. Data mining technique has substantial applications in the modern medical diagnostics. With the development of information technology and medical research, more effective models and algorithms will be applicable to support the disease diagnosis. |
