Deep Learning of Mesoscale Cortical Dynamics for Real-Time Classification of Forelimb Movement in Mice

Recent studies suggest that mesoscale cortical activity patterns encode movement-related information, such as direction and speed. We developed an approach to decode forelimb movement direction from mesoscale cortical dynamics in mice using wide-field GCaMP6f imaging. By simultaneously recording cortical activity and 3D kinematics of specific body points, we built a dataset linking 450 ms windows of cortical activity to movement categories (forward, backward, still). A convolutional neural network (CNN) trained on this dataset achieved real-time decoding at 100 Hz and sub- 10 ms latency, with a correlation of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{r}=0.72$</tex> between predicted and ground truth values. Although separate models were trained per animal, output distributions were highly consistent (mean inter-animal <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$r=0.91$</tex>), suggesting that mesoscale cortical dynamics related to movement are consistent across individuals. These results demonstrate the feasibility of the online decoding of movement types from wide-field cortical signals, towards minimally invasive upper limb neuroprosthetic.