ETHNOPRED: a novel machine learning method for accurate continental and sub-continental ancestry identification and population stratification correction

Table 7 Summary of the sub-continental classification problems results

Sub-continental problem	Number of subjects, split	Number of SNPs	Baseline	DT1 (Number of SNPs), Accuracy	Minimal Number of DTs (Number of SNPs), Accuracy	Number of Robust DTs (Number of SNPs)
European	267,	882895	61.8%	1 (10), 79.0% ± 5.6%	3 (31), 86.6% ± 2.4%	15 (180)
	CEU: 165
	TSI: 102
East Asian	250,	892833	54.8%	1 (12), 74.4% ± 7.9%	39 (502), 95.6% ± 3.9%	67 (877)
	CHB: 137
	JPT: 113
African	497,	616597	40.8%	1 (23), 66.2% ± 5.3%	21 (526), 95.6% ± 2.1%	157 (4236)
	LWK:110
	MKK: 184
	YRI: 203
North American	548,	526394	30.1%	1 (19), 82.7% ± 5.4%	11 (242), 98.4% ± 2.0%	70 (1643)
	ASW: 87
	CEU: 165
	CHD: 109
	GIH: 101
	MXL: 86
Kenyan	294,	781061	62.6%	1 (11), 79.2% ± 3.5%	25 (271), 95.9% ± 1.5%	31 (341)
	LWK: 110
	MKK: 184
Chinese	246,	829364	55.7%	1 (15), 47.2% ± 9.1%	- (−), ≤55.7%	- (−)
Chinese	CHB: 137	829364	55.7%	1 (15), 47.2% ± 9.1%	- (−), ≤55.7%	- (−)
	CHD: 109

This table summarizes the result of our studies on various sub-continental classification problems. The “Number of Subjects, Split” column shows the total number of subjects, followed by the list of (ethnic-group; number) pairs, giving the name of each subgroups and its size here. The “Number of SNPs” column gives the number of SNPs used for this study. The “Baseline” column gives the baseline accuracy of just using the majority class. The “DT1 (Number of SNPs), Accuracy” column provides the number of SNPs in the first decision tree, and its estimated 10-fold cross-validation accuracy. The “Minimal Number of DTs (Number of SNPs), Accuracy” column gives the minimal number of disjoint decision trees required to achieve the highest accuracy, and the number of SNPs involved, in these trees. The “Number of Robust DTs (Number of SNPs)” column gives the number of decision trees required to achieve robustness and the number of SNPs involved.

ISSN: 1471-2105