Neural Signal Decoding Using Machine Learning

Abstract

Brain-computer interfaces (BCIs) translate neural activity into control signals for prosthetic limbs, computer cursors, speech synthesisers, and communication devices. The decoding algorithm -- the mathematical model that maps neural signals to intended actions -- determines BCI performance. Traditional decoders (Kalman filters, population vector algorithms, Wiener filters) assume linear relationships between neural firing rates and movement kinematics. Deep learning decoders relax this assumption, learning nonlinear mappings from raw neural signals to high-dimensional behavioural outputs. We present the Neural Decoding Assessment Framework (NDAF), evaluating five ML decoding architectures -- recurrent neural networks (RNNs/LSTMs), temporal convolutional networks (TCNs), transformer-based sequence models, graph neural networks for electrode arrays, and hybrid Kalman-neural decoders -- across four neural signal modalities (intracortical microelectrode arrays, electrocorticography, stereoelectroencephalography, and scalp EEG). Our Neural Decoding Performance Score (NDPS) measures decoding accuracy, latency, long-term stability, cross-session generalisation, and computational efficiency for real-time deployment. Transformer-based decoders achieve the highest NDPS (0.924) through self-attention mechanisms that capture long-range temporal dependencies in neural population dynamics, while hybrid Kalman-neural decoders achieve the highest long-term stability (0.960) by combining physics-based state estimation with learned nonlinear observation models.