Table 1.

Comparison of the absolute performance of the ANN training methods in the simulation studies.

		No regularization	Noise injection	Weight decay	BANNa	Early stopping
50 training cases	Average AUC [95% CI]	0.723 [0.719, 0.727]	0.756 [0.751, 0.758]	0.742 [0.738, 0.746]		0.740 [0.737, 0.744]
	AUC standard deviation [95% CI]	0.043 [0.038, 0.048]	0.037 [0.033, 0.041]	0.050 [0.045, 0.055]		0.041 [0.035, 0.046]
50 training cases, complex ANNsb	Average AUC [95% CI]	0.694 [0.685, 0.703]	0.758 [0.755, 0.761]	0.745 [0.735, 0.754]		0.748 [0.740, 0.757]
	AUC standard deviation [95% CI]	0.044 [0.036, 0.052]	0.034 [0.031, 0.038]	0.048 [0.037, 0.060]		0.043 [0.032, 0.053]
100 training cases	Average AUC [95% CI]	0.762 [0.760, 0.765]	0.785 [0.784, 0.787]	0.784 [0.782, 0.786]		0.770 [0.768, 0.772]
	AUC standard deviation [95% CI]	0.028 [0.026, 0.030]	0.017 [0.016, 0.019]	0.020 [0.019, 0.022]		0.023 [0.021, 0.025]
200 training cases	Average AUC [95% CI]	0.788 [0.785, 0.790]	0.799 [0.797, 0.801]	0.797 [0.795, 0.799]	0.800 [0.798, 0.802]	0.790 [0.788, 0.792]
	AUC standard deviation [95% CI]	0.012 [0.011, 0.014]	0.009 [0.008, 0.010]	0.009 [0.007, 0.011]	0.009 [0.007, 0.010]	0.010 [0.008, 0.011]

^aThe BANNs failed to converge for training datasets with 50 and 100 cases.

^bThese ANNs had 20 hidden nodes and were trained to 1500 training iterations, whereas all other ANNs had 6 hidden nodes and were trained to 500 training iterations. The results were calculated at the 1485th and 485th training iteration, respectively.