Figure 1.
Schematic of the decoding algorithm.
Sub-band filtered multichannel LFP were processed using spatial filters to decode movement target direction. Common spatial patterns (CSP) filters were constructed to contrast binary classes of target directions, as well as of groups of directions. The main features of the projected data from each spatial filter were fed to their related Fisher linear discriminant (FLD) classifier. Finally, the outputs from all FLDs were processed using the error correcting output codes (ECOC) for the multi-class decision.
Figure 2.
Modulation of LFP across target direction.
(A) Time-frequency maps of LFP signals from one channel in M1 for 4 target directions (out of 8) during the instructed-delay center-out task. The data were aligned to the onset of movement (t = 0 s). The color map represents the change in power relative to baseline. The central arrows indicate the target direction associated with the different maps. The most reactive frequency bands were selected from these time-frequency maps for decomposing the LFP data into sub-bands. (B) Average LFP data in the 0.3–4 Hz band of two representative channels for the 90 deg and 270 deg target directions. The data were aligned to the onset of movement (t = 0 s). Note the difference in correlation between channels for these two directions. This characteristic was used for target direction decoding on a single trial basis.
Figure 3.
(A) Projections of 0.3–4 Hz LFP data from all channels using a spatial filter tuned to maximize the variance for the 90 deg direction and minimize the variance for the 270 deg direction. The projections were ordered along the ordinate according to target direction. The vertical dashed line indicates movement onset. (B) Variance of the spatial projection for each target direction in Fig. 3A. The variance was normalized to the maximum value. Note that the variance was maximum for 90 deg and minimum for 270 deg by design of the spatial filter, and that it varied progressively across direction. This characteristic indicates that LFP correlation structure across channels changed progressively with direction.
Figure 4.
Decoding accuracy and confusion matrices.
(A) Maximum circular correlation between predicted and actual target direction for each sub-band, brain area and subject. The results correspond to the cross-validated maximum correlation obtained during a trial. (B) Confusion matrices of decoding results using the 0.3–4 Hz and 48–200 Hz frequency bands combination from M1 and PMd data for each subject. A perfect prediction would result in all classifications being along the diagonal. Note that misclassifications occurred typically in neighboring directions evidencing that LFP for neighboring directions had similar correlation structures.
Figure 5.
Temporal evolution of decoding accuracy.
(A) Time-varying circular correlation between predicted and actual target direction for the two most informative sub-bands, and for each brain area and subject. The results at each time point were obtained using data from a time-window (500 ms for the cue and 1000 ms for movement onset aligned data, respectively) that preceded that time point. (B) Time-varying rate of change of circular correlation between predicted and actual target direction. Same conventions as in A. (C) Time-varying circular correlation between predicted and actual target direction using combined features obtained with the delta and upper gamma sub-bands from PMd and M1 for each subject. The thinner lines indicate the 95% confidence interval of the mean computed on the basis of the corrected t-test [25]; [ 26].
Table 1.
Decoding accuracy (DP: decoding power; ρT: circular correlation) across recording sessions.
Table 2.
Decoding accuracy (ρT) with SUA and LFP in single sessions.
Figure 6.
Decoding accuracy varies with number of channels and trials.
(A) Circular correlation between predicted and actual target direction computed as a function of number of channels used for decoding. (B) Circular correlation between predicted and actual target direction computed as a function of number of trials used for training the decoding algorithm (data of H564 in session 2).