Bayesian optimization of cortical neuroprosthetic vision using perceptual feedback
The challenge in cortical neuroprosthetic vision is determining the optimal, safe stimulation patterns for the visual cortex in order to evoke the desired perception in blind individuals--specifically, light perceptions known as phosphenes. Currently, clinical studies gain insights into the perceptual characteristics of the perceived phosphenes by asking for descriptions of provided stimulation protocols. However, the huge parameter space for multi-electrode stimulation settings makes it difficult to draw conclusions about the optimality of the stimulation patterns that lead to well-perceived phosphenes. A systematic search in the parameter space of the electrical stimulation is needed to achieve good perception. Bayesian optimization (BO) is a framework for finding optimal parameters efficiently. Using the patients scoring of the perception as feedback, a model of the patients response based on iteratively generated stimulation protocols can be built to maximize perception quality. A patient implanted with an intracortical 96-channel microelectrode array in their visual cortex was tested by iteratively presenting stimulation protocols, generated via BO for the first and random generation (RG) for the second experiment. Whereas standard BO methods do not scale well to problems with over a dozen inputs, we propose to optimize a set of 40 electrode currents using trust region-based BO. The generated protocols determine which electrodes are concurrently stimulated from the set and with how much current from a range of 0-50 {micro}A, on a maximum total current constraint of 500 {micro}A. The patient provided feedback for each stimulation based on their liking of the perception quality on a Likert scale, where a score of 7 indicated the highest quality and 0 no perception. In the BO experiment, the patient perception quality ratings gradually converged on higher values compared to the RG experiment. Similarly, gradually higher total current values were chosen by BO, in line with the observed preference of patients for higher currents due to brighter phosphenes. Finally, the electrodes that were observed to be more effective in producing phosphene perception in previous studies were gradually chosen more by BO also with the allocation of higher current values. This study demonstrates the power of BO in converging to optimal stimulation protocols based on patient feedback, providing a more efficient search for stimulation parameters for clinical studies.