Preservation of Partially Mixed Selectivity in Human Posterior Parietal Cortex across Changes in Task Context

Abstract

Recent studies in posterior parietal cortex (PPC) have found multiple effectors and cognitive strategies represented within a shared neural substrate in a structure termed "partially mixed selectivity" (Zhang et al., 2017). In this study, we examine whether the structure of these representations is preserved across changes in task context and is thus a robust and generalizable property of the neural population. Specifically, we test whether the structure is conserved from an open-loop motor imagery task (training) to a closed-loop cortical control task (online), a change that has led to substantial changes in neural behavior in prior studies in motor cortex. Recording from a 4 × 4 mm electrode array implanted in PPC of a human tetraplegic patient participating in a brain-machine interface (BMI) clinical trial, we studied the representations of imagined/attempted movements of the left/right hand and compare their individual BMI control performance using a one-dimensional cursor control task. We found that the structure of the representations is largely maintained between training and online control. Our results demonstrate for the first time that the structure observed in the context of an open-loop motor imagery task is maintained and accessible in the context of closed-loop BMI control. These results indicate that it is possible to decode the mixed variables found from a small patch of cortex in PPC and use them individually for BMI control. Furthermore, they show that the structure of the mixed representations is maintained and robust across changes in task context.

Keywords: bimanual; context; imagery; movement; neural prosthetic; parietal cortex.

Figure 1.

Experimental paradigm. A, Training task. The small red circle is the cursor, the gray circles are the possible targets, and the yellow circle is the target for the specific trial. B, Online control task.

Figure 2.

Tuning of the population to the conditions. A, Percentage of units tuned to each movement condition (bootstrap 95% CI, p < 0.05, uncorrected). B, Top row, Degree of specificity showing the distribution of how much units exclusively code ILH or ALH. Distribution during training shown in blue and distribution during online control shown in orange. For each distribution, the median and the probability the median is different from 0 (two-sided sign test) are shown in their corresponding colors. Middle row, Paired point plot showing how condition preferences for individual units changed from training to online control. Distribution during training and online control shown as a violin plot. For each distribution, the median and the probability the median is different from 0 (two-sided sign test) are shown underneath their corresponding x-label. Bottom row, Distributions showing change in preference values showing in the middle row between training and online control while preserving unit identity (fiducial, blue) and when shuffling unit identity (shuffled, gray). For each fiducial distribution, the median and the probability the median is different from 0 (two-sided sign test) are shown underneath their corresponding x-label. C, Similar to B, but for IRH and ARH. D, Similar to B but for ALH and ARH. E, Similar to B but for IRH and ILH. F, Correlation between movement representations during training and online control (bootstrap 95% CI).

Figure 3.

Possible configurations of representations and corresponding expected analysis results. A–C, Schematics for different possibilities in how the structure of the representations compares between training and online control. A, Schematic for the “structure maintained” case where the structure is consistent between training (left) and online control (right). Representations of the four movement conditions are separable during both training and online control, and in the same structure (i.e., the same configuration, as represented by the consistent placement of the conditions). B, Schematic for the “structure different” case where the movement conditions are separable during both training (left) and online control (right) but with different structures (i.e., different configurations). C, Schematic for the “structure collapsed” case where the movement conditions are separable during training only (left) and collapse into a single representation (as represented by the conditions being no longer separable in the online control case, right). D–F, Ideal expected result from cross-decoding analyses if the data follow the different schematics in Figure 3A–C. See Results for detailed explanation of colors and bars. Red lines represent chance performance (0.25). Performances significantly above chance are marked with an asterisk. D, Ideal expected result in the “structure maintained” case of Figure 3A. E, Ideal expected result in the structure different case. F, Ideal expected result in the structure collapsed case.

Figure 4.

Maintenance of the structure of representations. A, Results of the cross-decoding analysis performed on our data, presented as in Figure 3D–F. Performances significantly above chance are marked with an asterisk (one-sided Wilcoxon signed rank test, p < 0.05; see Materials and Methods for more details). B, Confusion matrices showing classifier predictions when generalizing from one context to the other, shown as the percentage of trials per condition. Columns are the true condition labels, and rows are the predicted labels. Left matrix corresponds to the classifier trained on the online control data and tested on the training data (Fig. 4A, left red bar). Right matrix corresponds to the classifier trained on the training data and tested on the online control data (Fig. 4A, right blue bar). C, Correlation between neural representations of pairs of runs where the runs were adjacent in time and matched in condition (blue), compared with the correlation between pairs adjacent in time mismatched in condition (red). Error bars are 95% bootstrapped confidence intervals. See Materials and Methods for more details. D, Example set of runs from a single session for the primary task paradigm (see Materials and Methods). Pairs marked in blue are matched by condition and are adjacent in time while pairs marked in red are mismatched in condition but still adjacent in time.

Figure 5.

Maintenance of structure of representations preserved for altered timing of condition blocks. A, Results of the cross-decoding analysis performed on additional data in which training and online test runs were collected with modified ordering of task conditions, presented as in Figure 4. Performances significantly above chance are marked with an asterisk (one-sided Wilcoxon signed rank test, p < 0.05; see Materials and Methods for more details). B, Confusion matrices showing classifier predictions when generalizing from one context to the other, shown as the percentage of trials per condition. Columns are the true condition labels, and rows are the predicted labels. Left matrix corresponds to the classifier trained on the online control data and tested on the training data (Fig. 4A, left red bar). Right matrix corresponds to the classifier trained on the training data and tested on the online control data (Fig. 4A, right blue bar). C, Correlation between neural representations of pairs of runs matched by condition but farther apart in time (blue) compared to pairs mismatched by condition but closer together in time (red). Error bars are 95% bootstrapped confidence intervals. See Materials and Methods for more details. D, Example of two blocks of runs (four runs per block) for the secondary task paradigm used to control for an order effect (see Materials and Methods). Pairs marked in blue are matched by condition and are close in time, pairs in yellow are matched by condition but are farther apart in time, while pairs in red are mismatched by condition but are closer in time.

Figure 6.

Maintenance of representations split by tuning preference. A, Average single-unit performance (weighted by the corresponding decoder weights) for imagined/attempted left-handed movements (bootstrap 95% CI). Units are grouped by tuning only to attempted movements, tuning only to imagined movements, and tuning to both. Performance was evaluated for imagined left-hand movements (blue bars) and attempted left-hand movements (red bars). Performances significantly above chance (one-sided sign test, p < 0.05, FDR corrected) are marked with an asterisk, and chance performance is marked by the dashed line. B, Similar to A but for right-handed movements. C, Average single-unit performance (weighted by the corresponding decoder weights) for left/right-handed movements using the attempt strategy. Units are grouped by specificity of tuning to the left or right hand, with performance evaluated during left- and right-handed movements (blue and red bars, respectively). Significant performances are marked. D, Similar to C but for movements using the imagine strategy.

Figure 7.

Online control performance. A, Performance of each movement condition, measured as the fraction of successful trials (bootstrap 95% CI). Dashed line indicates simulated chance performance (see Materials and Methods). B, Performance of each movement condition, measured as the mean duration of successful trials (bootstrap 95% CI). C, Mean R² of units tuned to each movement condition from Figure 2A (bootstrap 95% CI). D, Cross-validated R² of the decoder used for online control, trained on the training data for each condition (bootstrap 95% CI). Cross-validated R² was computed for each condition and session separately.