Background: Chronic kidney disease (CKD) is a significant micro-vascular complication of diabetes. Most patients missed the time for optimal treatment under current practice of clinical diagnosis. Better prediction of the CKD would reduce the morbidity and mortality. Objectives: The study aims to extensively test the most promising statistical methods for predictive modelling with simulated data and to complete a study protocol for predicting CKD in patients with type 2 diabetes. Methods: The profiles of two previous studies were used to generate simulated datasets for testing in this study. Forward selection, performance selection and full model approaches were used to select predictors for inclusion in prediction models. Predictive modelling with logistic regression and Cox proportional hazard regression methods were tested by cross-validation with the Hosmer-Lemeshow test for calibration, C statistics for discrimination, and net reclassification improvement and integrated discrimination improvement for reclassification. Results: Some predictor selection strategies led to unstable models with different selected variables, especially in the forward selection approach. All three predictor selection methods were consistent in performance of predictive modelling. Logistic regression models performed well in calibration while Cox proportional hazard regression models did not. The discrimination of prediction models was related to the incidence of CKD in the dataset, i.e., low prevalence achieved better discrimination. Increases in sample sizes improved the model stability, but not the model performance. Conclusion: Both development and validation of predictive models were conducted with simulated datasets to test the method feasibility of the most promising predictive modelling. A study protocol for predicting the risk of CKD in patients with type 2 diabetes was completed.

背景：慢性腎病是糖尿病重要的微血管併發症，當患者出現明顯的臨床症狀時， 往往已錯失最佳的治療時機。臨床中更好地預測慢性腎病可以降低其發病率和死 亡率。 目的：本研究的目的是在模擬數據集上廣泛地測試構建預測模型以及驗證模型績 效時最有希望使用的統計分析方法，完成 2 型糖尿病人併發慢性腎病的風險預測 模型的研究方案。 研究方法：以現有研究中的2個數據集概況為樣本產生模擬數據集。分別使用前 向選擇、績效選擇和全變量模型的方法確定最終進入模型的預測因子，建立 Logistic回歸模型和Cox比例風險回歸模型，交叉驗證下通過Hosmer-Lemeshow檢 驗衡量模型的標度，接受者操作特徵曲線下的面積衡量模型的區分度，以及再分 類指標 (淨再分類提升和和綜合區分指數) 衡量新增變量對模型區分效力的提 升。 研究結果：一些預測因子選擇策略導致構建的模型不穩定，每一次建模都可能納 入不同的變量，尤其是前向選擇。三種預測因子選擇策略所構建出的模型績效無 明顯差異。Logistic回歸模型表現出較好的標度，而Cox比例風險回歸模型的標度 不佳。模型的區分度與數據集本身的發病率有關，發病率越高，區分度越差。樣 本量的增加可以提高模型的穩定性，但其對模型績效的影響無統計學意義。 結論：本研究通過模擬數據集實現建立預測模型和檢驗模型績效的全過程，驗證 了數據分析中最有希望選用的統計方法的可行性。同時完成了2型糖尿病人患慢 性腎病的風險預測模型的研究方案。 關鍵詞：預測模型；2 型糖尿病患者；慢性腎病；模型驗證；模擬研究