Fig 1.
The sequence derived from logistic map for x (0) = 0.15 and A = 3.882.
Fig 2.
Flowchart for the generation of orthogonal chaotic code.
Fig 3.
The auto-correlation of m-sequence code (top) and chaotic code (bottom). The generated chaotic code follows the correlation property which is necessary in code modulation.
Fig 4.
One-sided amplitude spectrum of the presented stimuli (blue: Spectrum of the m-sequence codes stimuli, red: Spectrum of the chaotic code stimuli).
Dashed lines separate Low, Medium and High frequencies. Low: frequencies from 0 to 10 HZ, Medium: frequency range between 10 to 30 Hz and High: frequencies above 30 Hz. It is obvious that compared to the m-sequence codes, the chaotic codes frequency components are less in Low and Medium frequencies and are more in High frequency range.
Fig 5.
Time domain representation of m-sequence and chaotic codes.
Left and right columns show the m-sequence and chaotic codes respectively. M1-M4 and Ch1-Ch4 are the 4 shifted versions of m-sequences and chaotic codes respectively. Each box (shown with pink color) represents temporal shift of 0.088 second (8 bits) ahead with respect to previous one.
Fig 6.
The stimuli presentation diagram for single session: A single session consisted of 10 trials presenting m-sequence or chaotic code visual stimuli.
Each trial had 18 consecutively presented epochs. Each epoch presented a single visual stimulus code. Each trial was followed by 2 second break time. As there were 4 m-sequence (M1 − M4) and 4 chaotic codes (Ch1 − Ch4), each subject had total 8 sessions of stimulus presentation (see Fig 7 and text for details).
Table 1.
The stimuli specifications for both m-sequence and chaotic codes.
Fig 7.
Time sequences of activities of m-sequence (A) and chaotic code (B) presentation sessions. Each subject was presented with eight sessions. Each session started with 10 second rest and EEG recording and consisted of 10 trials. Each trial consisted of 18 epoch of consecutive m-sequence or chaotic codes presented with 2 seconds break after each trial. At the end of each session subjective fatigue rate was evaluated. The order of presentation of eight sessions for each subject was random.
Fig 8.
EEG recording electrodes placement according to 10–20 system.
Four active g.laddy bird electrodes were placed in Oz, O1, O2 and Pz. A2 and Fpz were selected as the reference and ground electrodes respectively.
Fig 9.
Signal recording set up: Stimuli were selected from stimuli presenting PC and sent to stimulator for presentation to subject via LED screen.
NI DAQ was used to record the trigger pulse coming from g.USBAmp and optic sensor output and also the stimulator box. The data was sent to PC for further analysis.
Fig 10.
Schematic representation of using CCA for template generation and target identification.
Template generation included extraction of epochs for each target and averaging them to generate templates. Target identification included extraction of epochs for testing trial and calculation of the canonical correlation of generated templates from training stage and averaged epochs for generating the feature vector and finally the maximum value of feature vector were selected.
Fig 11.
Schematic of representation of using STB for building beamformers and target identification.
Building beamformers included extraction of epochs for all the targets and generation of the activation patterns for each target and in parallel calculation of covariance matrix of concatenated epochs. The beamformers were calculated from Eq 4. The target identification included multiplying the beamformers with concatenated channels of averaged epochs in testing trials for the generation of feature vector and selecting the maximum score.
Fig 12.
Grand average and cross-correlations of evoked responses to m-sequences.
The grand average responses to codes Mi (i = 1:4) is shown with waveforms of their corresponding standard errors are shown with dotted plots (top) and the auto-correlation of response
and its cross-correlation with the responses
is shown with the waveforms of
(bottom). The delay between responses could be decoded from cross-correlation waveforms where they are maximum.
Fig 13.
Grand average and cross-correlations of evoked responses to chaotic codes.
The grand average of responses to codes Chi (i = 1:4) is shown with waveforms of their corresponding standard errors are shown with dotted plots (top) and the auto-correlation of response
and its cross-correlation with the responses
is shown with the waveforms of
(bottom). The delay between responses could be decoded from cross-correlation waveforms values (as shown) where they were maximum.
Fig 14.
Accuracies of target identification for the m-sequence and chaotic codes obtained from 10-fold cross-validation with CCA and STB methods over stimulation time.
Time duration for each epoch was 0.344 seconds and the total stimulation time for all 18 epochs was 6.2 seconds. The accuracy increased over stimulation time in both the methods. The dashed line shows that the STB is faster than CCA in reaching 70% accuracy.
Table 2.
Accuracies of target identification results of the 10-fold cross-validation for a trial for m-sequence and chaotic code for all subjects.
Table 3.
Statistical results for accuracy values of paired t-test in the comparison between STB and CCA methods for m-sequence and chaotic codes.
Fig 15.
Averaged subjective fatigue scores of all m-sequence and chaotic codes of all the subjects.
The chaotic codes VAS score was significantly lower than the m-sequence codes,*p = 0.0005, n = 44.
Fig 16.
Subjective fatigue scores of individual m-sequence and chaotic codes.
There was significant difference between VAS score of chaotic codes Ch1 and Ch3 (for each code n = 44 and *p = 0.002).