5.5 Real-time evaluation
6.3.3 Pattern classification
6.3.4.1 Pre-processing and feature extraction of deep brain LFPs
The deep brain LFP recordings were selected by excluding the segments contaminated with unintended movements based on surface EMGs. Also trials contaminated with artifacts, for which the behavioural response was incorrect, were removed using visual inspection. The LFPs were pre-processed with a low-pass Chebyshev Type I filter to remove the high frequencies. It was implemented with zero-phase shifting and a cut-off frequency of 90 Hz. A bandstop filter at 50 Hz to remove the power line noise was also implemented using a custom made adaptive filter (Wang et al. 2004). The recorded LFPs and their spectrogram (filtered at 90 Hz) from STN (a) and GPi (b) during an externally-cued left and right clicking movement tasks are presented in figure 6.2. The time of stimulus presentation and the subsequent motor response are shown using dotted and solid vertical lines, respectively. The state flow diagram for movement and its laterality decoding of deep brain (STN or GPi) LFPs and the general flowchart for decoding process of movement and laterality of LFPs presented in figure 6.3 (a) and (b) respectively.
Figure 6.2: Basal ganglia LFPs recorded from STN (a) for subject 1 and GPi (b) for subject 5, and their spectrogram (filtered at 90 Hz) during an externally-cued left or right clicking movement tasks. The time of stimulus presentation and the subsequent motor response are shown using dotted and solid vertical lines, respectively.
Figure 6.3: State flow diagram (a) for movement and its laterality decoding from deep brain (STN or GPi) LFPs and the general flowchart (b) for decoding process of movement and laterality of LFPs.
Figure 6.4: The flowchart for both instantaneous power and neural synchronisation feature extraction from the extracted LFP events.
The filtered LFPs were then re-sampled at 256 Hz and processed for feature extraction. The frequency dependent components were extracted as = 0-4 Hz, θ = 4-8 Hz, α = 8- 12 Hz, low β = 12-20 Hz, high β = 20-32 Hz, low γ = 32-60 Hz and high γ = 60-90 Hz frequency bands using a wavelet packet transform with a discrete Meyer wavelet (dmey) at decomposition scale 5. In each component of the LFPs, the event of left and right clicking tasks were segmented as 2 s before and after at each motor response registration. Similarly the resting activity was segmented as 2 s before and 2 s after at each stimulus registration. Segmented signal features were extracted based on instantaneous power and neural synchronisation from each frequency band. The flowchart for both feature extraction approach are presented in figure 6.4. To compute the instantaneous power features, the envelope of each component was computed using the Hilbert transform. The amplitude modulations of each component during left and right clicking events over all trials recorded from left STN of subject 1 are presented in
figure 6.5. From this figure it can be seen that there seems an amplitude decrease in the β band and increases in the , θ, α and γ bands, most visible in the band. Based on the average event related de-synchronisation and synchronisation in these bands (figure 6.6), the instantaneous power features for classification were defined as the average amplitude within each of five consecutive 100-ms windows in each frequency band. The five windows for the resting condition ran from -750 to -250 ms before the stimulus and the five windows for the clicking events ran from -150 to 350 ms around the motor response (figure 6.6). In total thirty-five features (each consisting of two values, one for left STN or GPi LFPs and the other for right STN or GPi LFPs) with five from each frequency band were obtained for classification.
Figure 6.5: The instantaneous amplitude of deep brain left STN LFPs components were computed using Hilbert transform and all trials for subject 1 are shown for the left and right clicking events with a 4-s window centred at the time of the response.
Figure 6.6: The average instantaneous amplitude of left STN LFPs in each components of subject 1 for the left and right clicking events with a 2-s window centred at the time of the response. The LFP signal features in each frequency bands were defined as average instantaneous amplitude within (W1:-150 to -50 ms), (W2:-50 to 50 ms), (W3:50 to 150 ms), (W4:150 to 250 ms) and (W5:250 to 350 ms) window around the response timing.
As it was assumed that synchronisation of the neural activities between different regions of the brain usually relates to the state or specific movements for which the signal is recorded, and will provide effective information about movements. The investigation of the dynamic changing of causal relationships between the neural signals recorded from the left and right STN or GPi for the events can provide more discriminative information to decode movement laterality. Therefore, causal strength between LFPs of left and right STN or GPi was evaluated by analysing Granger causality and denoted as the neural synchronisation feature for decoding left and right clicking events. During the causal analysis, the features were defined by computing contralateral and ipsilateral causal strength in each frequency band LFPs for each of the left and right clicking events. The contralateral and ipsilateral causal strength for left clicking events were computed as causality of right STN(GPi) → left STN(GPi) and left STN(GPi) → right STN(GPi) respectively. Similarly for right clicking events, the contralateral and
ipsilateral causal strength were computed as causality of left STN(GPi) → right STN(GPi) and right STN(GPi) → left STN(GPi) respectively. The analysis of Granger causality for each event was performed with segmented LFPs by varying analysis windows, 1) between 1 s before and after, and 2) between 500 ms before and after the onset of each motor response registration. For each analysis window, the MVAR model was estimated and the optimal order for the MVAR model was identified by locating the minimum of the AIC. However, the AIC dropped monotonically with increasing model order up to a value of 10 (5 and 15 in some cases) and then with the increase of the model order no further substantial decreases or increases of AIC was shown. Therefore average discriminability of contralateral and ipsilateral causal strength between left and right events produced from the pilot subjects and compared using model orders of 5, 8, 10, 12, 15, 20, 25 and 30, and observed that overall results varies. However, the model order 5, 10 and 25 produces more consistent and better discriminability in both analysis windows. Finally, a model order of 10 (40 ms) using a shorter window (500 ms before and after the onset) was selected as a tradeoff between sufficient discriminability and over-parameterisation (Brovelli et al. 2004). The analysed LFP data from all trials were treated as realisations of a common stochastic process, and thus were used to estimate the model coefficients for that process. Figure 6.7 presented the demonstration of contralateral and ipsilateral Granger causal strength between left and right STN for all trials and their average in all frequency bands of subject 1 for left (a) and right (b) events. Finally, fourteen neural synchronisation features were obtained for classification by evaluating contralateral and ipsilateral causal strength in each of seven frequency bands for left and right clicking events. The Granger causality analysis for extracting the neural synchronisation features was performed with the help of GCCA MATLAB toolbox (Seth 2010).
(a)
(b)
Figure 6.7: The contralateral and ipsilateral Granger causal strength between left and right STN for all trials and their average in all frequency bands of subject 1 for left (a) and right (b) events. The causal strength was defined for contralateral left event: right→left STN, right event: left→right STN, and for ipsilateral left event: left→right STN, right event: right→left STN in each frequency bands for each of left and right clicking events. The causal strength for each event was computed with in the 1-s
window (-500 to +500 ms) around the response timing in each frequency bands and defined as contralateral (red) and ipsilateral (black) neural synchronisation features.