Table 3.
Performance of the new and Korean undiagnosed diabetes screening method in the development and validation datasets.
| Model | Screeing method | Feature | AUC (95% CI) |
Youden index | Sensitivity (%) | Specificity (%) | PPV | NPV | PLR | NLR | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Train & Internal Validation Set | Lee model* | Risk score | Sex, Age, WC, Family history of diabetes, Hypertension status, Smoking status, Alcohol consumption |
0.750 (0.722 to 0.778) |
36 | 86 | 51 | 0.07 | 0.99 | 1.74 | 0.28 |
| Logistic Regression | Logistic Regression |
0.786 (0.761 to 0.811) |
42.1 | 89.50 | 52.60 | 0.08 | 0.99 | 1.88 | 0.2 | ||
| Random Forest | Random Forest Classifier |
0.781 (0.756 to 0.806) |
43.5 | 82.70 | 60.80 | 0.08 | 0.98 | 2021 | 0.22 | ||
| LGBM | LightGBM Classifier |
0.777 (0.751 to 0.803) |
42.4 | 80.80 | 61.50 | 0.08 | 0.98 | 2.26 | 0.21 | ||
| XGB | XGBoost Classifier |
0.786 (0.761 to 0.811) |
42.7 | 82.80 | 61.20 | 0.08 | 0.98 | 2.31 | 0.18 | ||
| Ada | AdaBoost Classifier |
0.785 (0.76 to 0.81) |
42.4 | 80.30 | 62.10 | 0.08 | 0.99 | 2.12 | 0.32 | ||
| External Validation set | Lee | Risk score | Sex, Age, WC, Family history of diabetes, Hypertension status, Smoking status, Alcohol consumption |
0.759 (0.741 to 0.777) |
36 | 90 | 46 | 0.08 | 0.99 | 1.67 | 0.21 |
| Logistic Regression | Logistic Regression |
0.801 (0.786 to 0.816) |
46.4 | 86.40 | 60.00 | 0.1 | 0.99 | 2.16 | 0.23 | ||
| Random Forest | Random Forest Classifier |
0.792 (0.776 to 0.808) |
46.1 | 83.00 | 63.10 | 0.11 | 0.99 | 2.25 | 0.27 | ||
| LGBM | LightGBM Classifier |
0.795 (0.779 to 0.811) |
45.8 | 81.90 | 64.00 | 0.11 | 0.98 | 2.27 | 0.28 | ||
| XGB | XGBoost Classifier |
0.802 (0.787 to 0.817) |
44.4 | 90.00 | 54.50 | 0.1 | 0.99 | 1.98 | 0.18 | ||
| Ada | AdaBoost Classifier |
0.784 (0.768 to 0.8) |
42.4 | 82.90 | 59.50 | 0.1 | 0.99 | 2.05 | 0.29 |
*Lee et al. 20125, When Lee model’s performance was tested, data from 2019, 2020 were used to build prediction model and data from 2014, 2015, 2016, 2017, 2018 were used to validate. WC: Waist circumference, RHR: Resting heart rate, LGBM: Light Gradient Boosting Machine, XGB: Extreme Gradient Boosting, Ada: Ada Boost, AUC: The receiver operating characteristics curve under the curve.
For this study, five different machine learning classification algorithms were used to predict undiagnosed diabetes. Based on their performance assessed by AUC, results from the best performed machine learning classification was used.