IEEE Computational Intelligence Magazine - May 2021 - 40

4) Drowsiness
In our experiments, we did not estimate accuracy but the number of false detections for the drowsiness classification because
the drowsiness dataset has unbalanced numbers for each class.
The proposed MSNN made the smallest number of mistakes in
decision making for passive BCI [6]. In particular, the proposed
method detected a driver's mental fatigue, i.e., drowsiness, from
the EEG signals. Our proposed method predicted 31.10 incorrect trials from a total of 177 samples on average. Furthermore,
accurately detecting a drowsy state is one of the most important MSNN capabilities for practical use. Our proposed model
only made 5.38 mistakes out of 35 drowsy trials on average,
thus exhibiting the highest precision score.
5) Seizure
Finally, for the seizure detection task, we also estimated the
number of false detections by following [14]. The MSNN
incorrectly identified 5.35 seizures among 178 total test seizure
samples. Furthermore, our proposed network was the fastest for

detecting seizures, i.e., it exhibited the shortest latency
time10 (approximately 4.98 sec on average) among various
methods. In other words, our proposed method demonstrated the best performance even with the shortest latency
time. Additionally, the proposed model correctly identified
approximately 92% of the seizures within 4.98 sec. We do
not present the standard deviation values for this seizure
detection experiment because each test trial consisted of
different numbers of seizures.11
V. Analyses and Discussions

In this section, we analyzed our proposed network. We determined the feature response by estimating PSD values and relevance scores [42] to show the multi-scale learning benefits. We
also visualized learned weights and represented features of the
10
To the best of our knowledge, it is still challenging to detect seizure onset early in
time (i.e., before the onset or short latency after the onset) [12], [14], [37].
11
See 'Supplementary B: Additional Experiment' for more performance comparison
with other methods.

TABLE II Performance evaluations. The Method column denotes all used classification/detection methods including baselines and
the proposed method on the various datasets, GIST-MI [40], KU-SSVEP [8], SEED-VIG [36], and CHB-MIT [12] EEG dataset. Each cell
depicts the average performance and the standard deviation of all subjects (or trials for the SEED-VIG [36]). For classification
performance on the SSVEP dataset, we used different kernel sizes for EEGNet [22] and the proposed method. These values are
marked by † and ‡, respectively. Furthermore, we estimated the number of false detections for SEED-VIG [36] and CHB-MIT [12]
dataset in the practical standpoint. * and ** respectively denote p < 0.05 and p < 0.01 estimated by two-tailed Wilcoxon's signedrank test between each Method and the proposed MSNN.
GIST-MI [40]

KU-SSVEP [8]

SEED-VIG [36]

CLASSIFICATION ACCURACY
METHOD

MEAN±STD.

CHB-MIT [12]

NUMBER OF FALSE DETECTIONS
MEAN±STD.

FALSE POSITIVE
(DROWSY)

MEAN (MEAN
LATENCY)

CSP + LDA [1]

.66±.14**

-

-

-

-

FBCSP + LDA [11]

.68±.15**

-

-

-

-

BSSFO + SVM [13]

.72±.17**

-

-

-

-

CCA [9]

-

.94±.09*

-

-

-

PSD + SVM [36]

-

-

31.20±15.47

6.74

-

SHOEB AND GUTTAG [14]

-

-

-

-

5.35 (5.11)

SHALLOW CONVNET [3]

.63±.11**

.52±.20**

34.89±19.13

6.51

19.21** (8.48)

DEEP CONVNET [3]

.61±.07**

.96±.08**

41.31±21.04**

8.65

8.74** (7.52)

RSTNN [18]

.69±.12**

.65±.20**

39.84±22.56**

8.08

24.35** (9.31)

ESTCNN [26]

.67±.10**

.79±.17**

41.10±21.31**

8.71

6.41** (7.01)

†

EEGNET [4], [22]

.64±.07**

.93±.10

46.63±22.10**

11.26

5.40** (6.23)

PARALLEL CRN [17]

.79±.11**

.93±.10

40.51±20.08**

9.32

12.45** (8.41)

CASCADE CRN [17]

.72±.09**

.94±.08

42.24±19.84**

9.84

18.37** (7.54)

MSNN (PROPOSED)

.81±.12

.93±.08‡

31.10±17.29

5.38

5.35 (4.98)

TABLE III Performance evaluations on the subject-independent manner. The Method row denotes all used classification methods
including baselines and the proposed method on the KU-MI [8] dataset. Each cell depicts the average performance and the
standard deviation of all target subjects. * denotes evaluation performances taken from [27].

40

METHOD

POOLED
CSP [16]

RAY
ET AL. [15]

MR
FBCSP [16]

PARALLEL
CRN [17]

CASCADE
CRN [17]

SSFR [27]

MSNN
(PROPOSED)

ACCURACY

.66±.16*

.67±.16*

.69±.15*

.72±.16

.73±.15

.74±.16*

.74±.18

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2021
IEEE Computational Intelligence Magazine - May 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2021
