Table 5.

Overall RSN or SOZ identification results for the three approaches.

Approach	RSN or SOZ vs. noise				SOZ vs. non-SOZ (RSN or noise)				Key observations
	Accuracy	Precision	Sensitivity	Specificity	Accuracy	Precision	Sensitivity	Specificity
EPIK (this paper)	71.7%	73.1%	72%	73.7%	84.7%	74.1%	88.6%	81.9%	Best performance for SOZ identification
LS-SVM (Hunyadi et al., 2014)	61.8%	52.2%	43%	73.6%	80.7%	52.2%	72.1%	78.7%	High false positives and false negatives Significant variance across patients
One sided t-test for +ve difference between EPIK and LS-SVM	p-value = ~0 [5, 15.2]	p-value = ~0 [18.7, 27.4]	p-value = ~0 [27.4, 45.9]	Rejected p-value = 0.9	p-value = ~0 [2, 6.5]	p-value = ~0 [20.7, 29.1]	p-value = ~0 [14.1 25.3]	Rejected p-value = 0.06
CNN (Nozais et al., 2021)	82.45%	82.7%	82.1%	81.5%	73.5%	28.5%	97.7%	42.85%	Best RSN identification performance. Poor SOZ performance due to lack of hand sorted SOZ IC examples.
One sided t-test for +ve difference between EPIK and CNN	Negative change P-value ~ 0 [−5.1, −13.2]	Rejected P-value = 0.6	Negative change P-value ~0 [−7, −12.1]	Negative change P-value = 0.02 [−4.1, −9]	P-value ~0 [8.3, 15.7]	P-value ~ 0 [51.2, 60]	Negative change P-value ~ 0 [−4.2, −11.3]	P-value = 0.001 [31.6, 45.2]