An EMG foundation model for neural decoding

Decoding algorithms can be used to predict motor behaviour from patterns of neural activity. However, most studies rely on subject-optimized models, limiting generalization and scalability to novel subjects and tasks. Building on recent advances in deep learning and large-scale data, here we developed an EMG foundation model for neural decoding. Our model was trained on more than 197 hours of neural recordings from 1,667 subjects. We used unsupervised learning to pretrain our encoder layers on unlabeled data, followed by supervised learning on our benchmark dataset. Additionally, we performed large-scale architecture searches to develop a custom encoder-decoder model composed of convolutional and transformer layers, optimized for both scalability and performance. Our foundation model consistently outperformed the previous state-of-the-art (i.e., subject-optimized models) across both in-distribution and out-of-distribution evaluations. For in-distribution evaluation, few-shot fine-tuning yielded an average F1 score of 0.697, compared to 0.638 for subject-optimized models. For out-of-distribution evaluation on clinical and demographically-shifted subjects, we achieved an average F1 score of 0.599, compared to 0.518 for the subject-optimized baselines. Taken together, our results highlight the value of foundation models for robust and generalizable neural decoding. By publicly releasing our neural network weights and training pipeline, we aim to support future research in computational neuroscience and neural-machine interfaces.