Table 1 Comparison of the crossNN model to ad-hoc RFs²³ and the Sturgeon DNN approach²⁵

Cohort	Number of cases	Metric	crossNN	Sturgeon 0.8	Sturgeon 0.95	Ad-hoc RF
450K	610	Accuracy	0.979	0.962	0.962	0.97
		Precision	0.996	0.973	0.988	0.972
		Sensitivity	0.93	0.861	0.792	0.966
		AUC	0.973	0.892	0.892	0.921
EPICv1	554	Accuracy	0.948	0.955	0.955	0.966
		Precision	0.99	0.963	0.967	0.971
		Sensitivity	0.894	0.944	0.908	0.96
		AUC	0.953	0.773	0.773	0.884
EPICv2	133	Accuracy	0.97	1	1	0.985
		Precision	1	1	1	0.992
		Sensitivity	0.895	0.977	0.94	0.985
		AUC	0.986	NaN	NaN	0.992
Nanopore R9	415	Accuracy	0.964	0.925	0.925	0.937
		Precision	0.99	0.964	0.973	0.99
		Sensitivity	0.908	0.824	0.61	0.718
		AUC	0.967	0.843	0.843	0.917
Nanopore R10	129	Accuracy	0.922	0.884	0.884	0.899
		precision	0.965	0.954	0.987	1
		sensitivity	0.853	0.791	0.581	0.674
		AUC	0.931	0.905	0.905	0.914
Targeted sequencing	124	Accuracy	0.895	0.855	0.855	0.839
		Precision	0.991	0.994	1	0.99
		Sensitivity	0.879	0.806	0.726	0.766
		AUC	0.997	0.954	0.954	0.958
WGBS	125	Accuracy	0.936	0.808	0.808	0.88
		Precision	0.991	0.892	0.922	0.979
		Sensitivity	0.848	0.616	0.432	0.736
		AUC	0.94	0.79	0.79	0.918
Overall	2,090	Accuracy	0.956	0.935	0.935	0.946
		Precision	0.991	0.963	0.978	0.947
		Sensitivity	0.901	0.861	0.757	0.873
		AUC	0.953	0.865	0.865	0.9

For each model, MCF-level raw accuracy before the application of cutoffs, precision with platform-specific cutoffs and AUC of the ROC for the (calibrated) score to predict the correct classification are given. For crossNN, the following cutoffs, as derived above, were used: microarray > 0.4; crossNN nanopore/targeted methyl-seq/WGBS > 0.2. Published validated cutoffs were used for ad-hoc RF and the Sturgeon DNN (ad-hoc RF > 0.15; Sturgeon DNN > 0.8 or > 0.95, respectively). NaN, not a number. Bold indicates row-wise maximum values.

Quick links

Search