Screening for Prediabetes Using Machine Learning Models

Table 2

Performance of the ANN, SVM, and screening score (Lee et al. [8]) models using the internal and external validation sets for predicting prediabetes.

		AUC	Accuracy (%)	Sensitivity (%)	Specificity (%)
Internal validation set; (𝑛 = 1,551)	ANN*	0.768	69.0	74.1	67.5
	SVM^†	0.761	64.9	78.9	61.2
	Screening score^‡	0.734	63.4	76.1	60.0
External validation set; (𝑛 = 4,566)	ANN*	0.729	60.7	77.2	56.7
	SVM^†	0.731	66.1	69.4	65.3
	Screening score^‡	0.712	59.9	74.3	56.4

AUC: area under the curve; ANN: artificial neural network; SVM: support vector machine.
The internal validation set was comprised of data from the Korean National Health and Nutrition Examination Survey (KNHANES) 2010, and the external validation set included data from KNHANES 2011. *The chosen model was a multilayer perceptron model with 1 hidden layer, batch training, and momentum learning (MLP-1-B-M) of backpropagation feedforward algorithm. ^†The optimal model was found using Gaussian kernel function with a penalty parameter (C) of 10 and scaling factor (σ) of 10. ^‡The performance was calculated by applying the screening score model for prediabetes based on that of Lee et al. [8] to the data from KNHANES 2010 and 2011.