TABLE 2.

Sex detection accuracies of per‐subject predictions from majority votes over 50 segments (mean ± SD over 30 randomly initialized networks).

Experiment a	Kernels b	Accuracy		p c	p d
		Balanced	Imbalanced	$\alpha /n=0.003$	$\alpha /n=0.01$
EEG
Raw, 1–40 Hz	21 × 19	$78\%\pm 2\%$	$78\%\pm 2\%$	$<.001$
ICA, 1–40 Hz	21 × 19	$74\%\pm 2\%$	$75\%\pm 2\%$	$<.001$
1–4 Hz $\left(\delta \right)$	21 × 19	$73\%\pm 2\%$	$74\%\pm 2\%$	$<.001$
4–8 Hz $\left(\theta \right)$	21 × 19	$73\%\pm 3\%$	$74\%\pm 3\%$	$<.001$
8–14 Hz $\left(\alpha \right)$	21 × 19	$74\%\pm 2\%$	$75\%\pm 2\%$	$<.001$
14–30 Hz $\left(\beta \right)$	21 × 19	$71\%\pm 3\%$	$72\%\pm 2\%$	$<.001$
30–40 Hz $\left(\gamma \right)$	21 × 19	$70\%\pm 3\%$	$70\%\pm 2\%$	$<.001$
ICA, 1–40 Hz, shuffled time points	21 × 1	$68\%\pm 3\%$	$69\%\pm 2\%$	$.002$	$\Lsh$
1–4 Hz $\left(\delta \right)$	21 × 1	$72\%\pm 2\%$	$73\%\pm 2\%$	$<.001$	$<.001$
4–8 Hz $\left(\theta \right)$	21 × 1	$71\%\pm 2\%$	$71\%\pm 2\%$	$<.001$	$0.007$
8–14 Hz $\left(\alpha \right)$	21 × 1	$73\%\pm 2\%$	$73\%\pm 2\%$	$<.001$	$<.001$
14–30 Hz $\left(\beta \right)$	21 × 1	$72\%\pm 2\%$	$73\%\pm 2\%$	$<.001$	$<.001$
30–40 Hz $\left(\gamma \right)$	21 × 1	$70\%\pm 2\%$	$70\%\pm 2\%$	$<.001$	$0.11$
ICA, 1–40 Hz random channel phase e	21 × 19	$61\%\pm 3\%$	$61\%\pm 3\%$	$.17$
Raw, 1–40 Hz, ensemble	21 × 19	81%	$82\%$	$<.001$
ECG
Raw, 1–40 Hz	1 × 19	$58\%\pm 2\%$	$60\%\pm 1\%$	$.22$

Note: To probe the accuracy limit of this network architecture, we increased the number of kernels from 16 to 512 and made an ensemble prediction by a majority vote from 30 independently randomly initialized networks, trained on the same data (bold).

^a Filters and modifications were applied to training and evaluation data.

^b Size of the convolutional kernels in the CNN; channels × samples; $19\text{samples}\widehat{=}76\mathrm{ms}$.

^c Binomial test on the (imbalanced) accuracy; 142 subjects, 57% female; Bonferroni adjusted alpha level of 0.003 per test ($0.05/15$) for $n=15$ multiple comparisons.

^d One‐sided paired t‐test on the (imbalanced) accuracy to compare the accuracies from the physiological frequency bands against the 1–40 Hz data; 142 subjects, 57% female.

^e Without fixed temporal relations between channels it is harder to populate multiple channels of one kernel. To neutralize this drawback, we increased the number of kernels for this experiment from 16 to 512.