Researchers from the University of California, Santa Barbara, ETH Zurich, and Miguel Hernández University have demonstrated that deep learning can significantly enhance the precision and adaptability of visual cortical prostheses. The study, published this week in the journal Neuron, outlines a proof-of-concept approach that uses artificial intelligence to optimize electrical stimulation patterns for brain implants, advancing the development of functional bionic vision systems. Visual prostheses traditionally target the retina or optic nerves, but cortical implants bypass damaged pathways to stimulate the visual cortex directly. This strategy is particularly relevant for patients with strokes, neurodegenerative diseases, or traumatic injuries who retain intact visual cortices. In 2024, the research team conducted clinical trials at Hospital IMED Elche in Spain with a 27-year-old participant who lost his sight due to trauma. The participant wore a temporary 96-channel electrode array implanted in his visual cortex, which elicited phosphenes, or perceived flashes of light. Rather than relying on static stimulation protocols, the researchers trained a deep neural network to predict cortical responses using varying electrical parameters and the brain's resting activity. By integrating real-time neural recordings, the model identified stimulation patterns that reproducibly triggered targeted brain activity while requiring lower electrical currents. The recorded neural responses proved more accurate predictors of the participant's visual perceptions than the stimulation settings alone, underscoring the complex, non-linear transformation between electrical input and conscious sight. The AI framework also incorporates dynamic adjustments for fluctuations in baseline brain activity, a critical requirement for long-term reliability. As neural states naturally vary, the device can continuously recalibrate its output to maintain consistent perceptual outcomes. This adaptive capability addresses a fundamental challenge in neuroprosthetics, shifting the field from rigid, pre-programmed hardware to personalized, learning-driven systems that evolve with the user. Co-led by researchers from ETH Zurich, UCSB, and Miguel Hernández University, the project was supervised by Professors Michael Beyeler, Shih-Chii Liu, and Eduardo Fernández. Beyeler noted that translating computational models into live human trials validated the system's practical utility. The findings represent a notable advancement in closed-loop neuromodulation, confirming that artificial intelligence can effectively decode and shape cortical interactions to deliver more predictable visual restoration. Subsequent phases of the Spanish feasibility trial will further refine the technology, with the objective of transitioning visual cortical prostheses from experimental prototypes to viable clinical treatments for severe vision loss.