TABLE 3.

Classification results of the SNN in iEEG, ECoG and scalp EEG.

Classification task and EEG modality	Seizure outcome prediction in patients implanted with iEEG. (HFO area resected →seizure freedom; HFO area not resected → seizure recurrence)			Seizure outcome prediction in patients implanted with ECoG. (max HFO rate ≥ 1 HFO/min → seizure recurrence; max HFO rate < 1 HFO/min → seizure freedom)		Active epilepsy prediction (seizures/month > 0) using scalp EEG recordings from pediatric patients (mean HFO rate > 0.25 HFO/min → active epilepsy; HFO rate < 0.25 HFO/min → seizure freedom)
SNN architecture used for HFO detection	Morphology detector	Core SNN in Brian 2	DYNAP-SE hardware	Spectrum detector	Core SNN + in-band artifact rejection SNN in Brian 2	Spectrum detector	Core SNN + in-band artifact rejection SNN + artifact detection SNN in Brian 2
Specificity = TN/(TN + FP)	100 (54 100%)	100 (54 100%)	100 (54 100%)	100 (63 100%)	100 (63 100%)	67 (57 98%)	100 (69 100%)
Sensitivity = TP/(TP + FN)	0 (0 71%)	33 (1 91%)	33 (1 91%)	100 (63 100%)	100 (63 100%)	86 (57 98%)	71 (42 92%)
Negative Predictive Value = TN/(TN + FN)	67 (30 93%)	75 (35 97%)	75 (35 97%)	100 (63 100%)	100 (63 100%)	67 (22 96%)	71 (42 92%)
Positive Predictive Value = TP/(TP + FP)	–	100 (3 100%)	100 (3 100%)	100 (63 100%)	100 (63 100%)	86 (57 98%)	100 (69 100%)
Accuracy = (TP + TN)/N (%)	67 (30 93%)	78 (40 97%)	78 (40 97%)	100 (63 100%)	100 (63 100%)	85 (62 97%)	80 (56 94%)

TP True Positive; TN True Negative; FP False Positive; FN False Negative; N = TP + TN + FP + FN = number of patients.

To find HFO in the iEEG dataset, we first used the SNN simulator Brian 2 and the toolbox Teili (Milde, 2018), to simulate our SNN and then we mapped the SNN in hardware using the neuromorphic processor DYNAP-SE (Moradi et al., 2018).

The numbers in brackets are the 95% confidence intervals (CI).