Statistical analysis

The present study’s primary outcome was HbA1c ≥9%. Categorical variables were classified by clinical findings, which were expressed as frequency and percentage. Continuous variables were expressed as mean and standard deviation. Descriptive statistics were used to analyze the difference between the development and the validation groups. The Chi-square test was used to compare the frequency distribution of categorical variables. The Shapiro-Wilk test and independent sample t-test were used for continuous variables of the normal distribution test and the difference between groups, respectively.

All the variables were evaluated using univariable logistic regression analysis. The model was initiated with 26 candidate variables and reduced to find the best-fitting model. Multivariable stepwise (backward) logistic regression analysis was conducted to assess factors associated with uncontrolled diabetes, using a significance level of p-value < 0.2 as the criterion to remove variables from the models. Odds ratios (OR) and 95% confidence intervals (95% CI) were calculated for significant variables. The Variance Inflation Factor (VIF) was employed to address multicollinearity among independent variables. Variables with statistical significance in both univariable and multivariable analyses were combined to predict uncontrolled diabetes.

The final prediction model was applied to develop an effective prognostic nomogram. The nomogram performance was evaluated regarding differentiation, calibration, and clinical validity. Receiver Operating Characteristic (ROC) curve analysis and the area under the curve (AUC) were used to measure nomogram accuracy. The AUC neared 1, the more the maximum AUC, and the perfection in the differentiation between the diseased and non-diseased improved. The true positive fraction (sensitivity), false-positive fraction (1-specificity), positive predictive value, negative predictive value, positive likelihood ratio, and negative likelihood ratio were established. Youden’s index maximizes the difference between TPF (sensitivity) and FPF (1-specificity), which calculates the optimal cut-off point on the ROC curve. The model correction used the Hosmer-Lemeshow goodness-of-fit test with a p-value greater than 0.05. No distinct difference was found between the predicted and actual values. Decision curve analysis was used to evaluate the clinical validity of the model. In addition, the ROC curve point defined the optimal cut-off value, which was selected using Youden’s index (sensitivity+specificity-1).