Table 2 Performance comparison of micro-average C-indexes for seven models on seven cancer types based on a different number of genes selected by CWx.

Model	6407	3000	2000	1000	500	196	144	100	81	49	25	Average
CNN-Cox	0.6441	0.6742	0.6860	0.6803	0.6816	0.6861	0.6797	0.6801	0.6770	0.6716	0.6704	0.6755
1D-CNNCox	0.6438	0.6739	0.6857	0.6813	0.6741	0.6859	0.6789	0.6771	0.6770	0.6682	0.6647	0.6737
NN-Cox	0.6288	0.6492	0.6625	0.6652	0.6606	0.6741	0.6686	0.6572	0.6553	0.6481	0.6485	0.6562
RSF	0.6468	0.6624	0.6683	0.6736	0.6722	0.6817	0.6789	0.6761	0.6774	0.6748	0.6703	0.6711
GBM	0.6254	0.6428	0.6557	0.6552	0.6742	0.6722	0.6683	0.6670	0.6527	0.6559	0.6501	0.6563
Cox-EN	0.6375	0.6901	0.6885	0.6744	0.6326	0.6589	0.6553	0.6589	0.6650	0.6661	0.6667	0.6631
SSVM	0.6383	0.6923	0.6939	0.6698	0.6111	0.6180	0.6113	0.6395	0.6514	0.6586	0.6675	0.6502

Bold values facilitate the rapid identification of the optimal performance survival analysis moder under different experimental conditions, that is, with different number of selected gene features.
Friedman test was performed on micro-average C-indexes of seven models on seven cancer types, BLCA, SKCM, KIRC, LGG, HNSC, LUAD, and LUSC. F_F = 9.7206 is far greater than the critical value 2.2541 at α = 0.05 significance level, showing seven models perform significantly differently. The post-hoc Bonferroni–Dunn test was conducted for paired comparisons of CNN-Cox against other baseline models. Critical difference (CD) diagram for test results is shown, where x axis denotes average ranks. If the rank difference between the two methods is smaller than CD = 2.490, performance difference is not significant (horizontal line). Except 1D-CNNCox and RSF, CNN-Cox has significantly better performance than other models.

Search