Classification of Thyroid Nodules with Stacked Denoising Sparse Autoencoder

Table 3

Performance of the IRC based on SDSAE.

Parameter	Training set (n = 929)		Independent testing set (n = 84)
	Value	CI of 95%	Value	CI of 95%
Proportion of malignancy	85.8%	—	85.7%	—
Area under the curve	0.887	0.855–0.920	0.785	0.638–0.931
Accuracy	92.0%	92.0–92.0%	92.9%	92.7–93.0%
Sensitivity	93.4%	91.6–95.1%	98.6%	95.9–101.3%
Specificity	84.1%	77.9–90.3%	58.3%	30.4–86.2%
Positive likelihood ratio	5.868	3.962–8.689	2.367	1.211–4.625
Negative likelihood ratio	0.079	0.060–0.104	0.024	0.003–0.177
Positive predictive value	97.3%	96.1–98.4%	93.4%	87.8–99%
Negative predictive value	67.7%	60.5–74.8%	87.5%	64.6–110.4%

IRC: immune-related classifier; SDSAE: stacked denoising sparse autoencoder; CI: confidence interval.