Dissecting the computational necessity of hybrid deep learning models for decoding complex motor behavior from neuronal population activity

Decoding complex motor behavior from multiscale brain networks remains a fundamental challenge in neuroscience. Advances in in vivo optical imaging now enable the recording of high-dimensional neuronal population activity at cellular resolution, necessitating decoding frameworks capable of capturing both spatial and temporal structure. Here, we employ an attention-based hybrid convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) architecture to decode skilled forelimb movements from in vivo two-photon calcium imaging of excitatory neuronal ensembles in unilateral M1. By jointly modeling spatial organization across neuronal populations and temporal structure in activity sequences, the hybrid framework enabled reliable decoding of both ipsilateral and contralateral forelimb movements during a skilled locomotor task. Ablation analyses further demonstrate that integrating spatial and temporal components is essential for maintaining decoding performance, highlighting the computational necessity of a hybrid architecture. Together, these results establish hybrid deep learning models as an effective framework for extracting behaviorally relevant spatiotemporal features from mesoscale neuronal network activity.