Dual-VCT: A dual-branch VMD-CNN-transformer model for local field potentials decoding

Abstract Objective. Local field potential (LFP) decoding is critical for the clinical translation of intracortical brain-machine interfaces, yet existing decoding methods are limited by three key bottlenecks: insufficient single-scale feature utilization, inefficient multi-scale feature fusion, and poor robustness across task paradigms and chronic recording conditions. Approach. To address these challenges, we propose Dual-VCT, a novel dual-branch variational mode decomposition-convolutional neural network-Transformer (VMD-CNN-Transformer) model for end-to-end LFP decoding. The core innovation of Dual-VCT is its symmetric time-frequency parallel architecture with independent VMD modules embedded in both branches: a temporal branch decomposes local motor potential (LMP) signals via VMD to capture motion-related instantaneous neural activity, while a frequency-domain branch leverages VMD to isolate task-relevant spectral power components, with a hierarchical fusion pipeline enabling robust cross-scale feature integration. Main results. Validated in non-human primate experiments, Dual-VCT achieved a classification accuracy of 0.930 ± 0.023 in the 3-class spatial grasping task, and a Pearson correlation coefficient (CC) of 0.910 ± 0.023 in the finger point-to-point tracking task. It significantly outperformed all comparative dual-branch methods under identical experimental conditions ( p < 0.05), delivered a 4% performance gain over single-feature decoding, and exhibited strong cross-task robustness and cross-day stability. Ablation experiments confirmed the core contribution of the dual-branch VMD design. Significance. This work provides a high-performance structured paradigm for LFP decoding, with a clinically oriented design that supports the long-term stability of chronic iBMI systems.