A deep learning framework was trained on 26-session intracortical recordings from a 96-channel Utah Electrode Array (UEA) implanted in the occipital cortex of one blind adult to control electrically evoked population activity in human visual cortex. Two strategies were compared—a gradient-based optimizer and a real-time inverse neural network—against conventional 1-to-1 mapping and linear baselines.
Both deep learning methods significantly outperformed all baselines in reproducing target neural response patterns (p<0.05). Including recorded neural activity (ΔMUAe + pre-stimulus MUAe) alongside stimulation parameters improved phosphene detection accuracy from 74.2% to 88.7% (56% error reduction), color prediction from 49.4% to 76.7% (54% error reduction), and brightness prediction from 44.8% to 60.9% (29% error reduction). Optimized stimuli required lower currents, and manifold distance to natural neural activity strongly predicted control error (r=0.85).
- Single participant (27-year-old blind male), so generalizability to other patients or implant sites is unknown. - The forward model does not account for long-term cortical plasticity or adaptive changes over months. - Gradient-based optimization takes ~10–20 seconds per stimulus, limiting real-time use.
For future cortical visual prostheses, monitoring evoked neural population activity—not just stimulation parameters—substantially improves prediction of what a patient will perceive. Deep learning-based stimulus synthesis can reduce required current amplitudes while producing more consistent phosphenes, pointing toward safer and more reliable closed-loop neuroprosthetic systems.
Explore related topics
