Polytomous diagnosis of ovarian tumors as benign, borderline, primary invasive or metastatic: development and validation of standard and kernel-based risk prediction models

Table 3 Validation results using a polytomous c-index

	Internal validation (n = 312)		Temporal validation (n = 941)		External validation (n = 997)
Model (# predictors)	Polytomous c-index (95% CI)	Difference with best model (95% CI)	Polytomous c-index (95% CI)	Difference with best model (95% CI)	Polytomous c-index (95% CI)	Difference with best model (95% CI)
Group 1: logistic regression models
LR-PC (10)	.67 (.58-.75)	-	.60 (.56-.65)	-	.60 (.55-.65)	-
MLR (9)	.64 (.56-.73)	.025 (-.004; .053)	.58 (.54-.62)	.020 (.000; .040)	.58 (.53-.62)	.028 (.000; .058)
Group 2: kernel-based and logistic regression models (based on R1U variable selection)*
LR-PC2 (11)	.69 (.60-.77)	-	.59 (.55-.64)	-	.64 (.59-.68)	-
KLR-PC (11)	.67 (.59-.75)	.016 (-.013; .051)	.58 (.54-.63)	.012 (-.006; .027)	.61 (.57-.66)	.026 (.004; .049)
LSSVM-PC (11)	.66 (.58-.75)	.025 (-.007; .060)	.58 (.54-.62)	.015 (-.005; .035)	.61 (.57-.65)	.028 (.005; .052)
MKLR (11)	.64 (.56-.73)	.046 (.003; .086)	.57 (.52-.62)	.027 (.000; .056)	.58 (.53-.62)	.060 (.033; .092)

Models are ranked by the value of the polytomous c-index. Ben: benign; PrInv: primary invasive; Bord: borderline; Meta: metastatic; CI: confidence interval. 95% CIs are computed using the bias-corrected bootstrap method using 1000 bootstrap samples.
* R1U variable selection: variable selection within the framework of least squares support vector machines that is based on rank 1 updates of the kernel matrix [27].

ISSN: 1471-2288