Deep Learning-Based Control of Electrically Evoked Activity in Human Visual Cortex.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Current visual cortical prostheses provide limited sight restoration due to crude percepts and inefficient manual calibration.
• Scalability and precision remain significant challenges in neuroprosthetic device control.

Purpose of the Study:

• To develop and validate an automated, data-driven neural control method for visual neuroprostheses using deep learning.
• To generate optimal multi-electrode stimulation patterns for evoking targeted neural responses and improving visual percept stability.

Main Methods:

• A deep neural network was trained on single-trial evoked responses from a 96-channel Utah electrode array in a blind participant's occipital cortex.
• Two control strategies were implemented: a learned inverse network for real-time synthesis and a gradient-based optimizer for precise neural response targeting.

Main Results:

• The automated method significantly outperformed conventional approaches in controlling neural activity and required lower stimulation currents.
• Stimulation parameters adapted to resting-state data, leading to more stable and reliable visual percepts.
• Neural population activity was a better predictor of perceptual outcomes than stimulation parameters alone.

Conclusions:

• Data-driven neural control is feasible in human implants for visual neuroprosthetics.
• This framework provides a foundation for next-generation, model-driven neuroprosthetic systems to enhance sensory restoration.
• The developed deep learning approach offers a scalable and precise method for visual neuroprosthetic control.