The study of an integrated diabetes prediction model based on user-defined risk decision making strategy

Diabetes, as a typical chronic disease, seriously affects the quality of life and economic burden of patients and their families. If the early and accurate prediction of diabetes can be realized according to patients' health-data, the development of diabetes would be effectively prevented or mitigated. The existing diabetes prediction models have achieved better prediction results. However, noise and high-dimensional nonlinear features in health-data may still be important factors affecting the performance of prediction models and deserve further research. Aiming at these, we propose an integrated diabetes prediction model based on user-defined risk decision making strategies, named FDNN. The model tandemly integrates modules for data-preprocessing, diabetes prediction models based on different features, and user-defined risk decision making strategies. The problems of data-standardization, data noise filtering, and high-dimensional nonlinear features learning are solved hierarchically in the model. Finally, based on the PIMA dataset, FDNN is compared with baselines such as SVM, Logistic, random forest, and gradient boosting decision tree (GBDT). The experimental results show that noise does exist in the diabetes dataset and the noise is an important factor affecting the prediction performance. FDNN with feature selection is proposed to mitigate the impact of noise in diabetes dataset on prediction performance. In the dataset with noise, the sensitivity, specificity, accuracy and area under the receiver operating characteristic curve of FDNN are 0.9278, 0.6667, 0.8312, and 0.8475 respectively, which are better than the comparison algorithms. It is 6.75%, 11.92%, 5.79%, and 0.83% better than GBDT respectively, with more accurate and stable prediction performance.