Recalibration of neuromodulation parameters in neural implants with adaptive Bayesian optimization

Abstract

Objective. Neuromodulation technology holds promise for treating conditions where physiological mechanisms of neural activity have been affected. To make treatments efficient and devices highly effective, neurostimulation protocols must be personalized. The interface between the targeted nervous tissue and the neurotechnology (i.e. human-machine link or neural interface) usually requires constant re-calibration of neuromodulation parameters, due to many different biological and microscale phenomena happening over-time. This adaptation of the optimal stimulation parameters generally involves an expert-mediated re-calibration, with corresponding economic burden, compromised every-day usability and efficacy of the device, and consequent loss of time and increased discomfort of patients going back to clinics to get the device tuned. We aim to construct an adaptable AI-based system, able to compensate for these changes autonomously. Approach. We exploited Gaussian process-based Bayesian optimization (GPBO) methods to re-adjust the neurostimulation parameters in realistic neuroprosthetic data by integrating temporal information into the process to tackle the issue of time variability. To this aim, we built a predictive model able to tune the neuromodulation parameters in two separate crucial scenarios where re-calibration is needed. In the first one, we built a model able to find the optimal active sites in a multichannel electrode, i.e. able to cover a certain function for a neuroprosthesis, which in this specific case was the evoked-sensation location variability. In the second one, we propose an algorithm able to adapt the injected charge required to obtain a functional neural activation (e.g. perceptual threshold variability). By retrospectively collecting the outcomes from the calibration experiments in a human clinical trial utilizing implantable neuromodulation devices, we were able to quantitatively assess our GPBO-based approach in an offline setting. Main results. Our automatic algorithm can successfully adapt neurostimulation parameters to evoked-sensation location changes and to perceptual threshold changes over-time. These findings propose a quick, automatic way to tackle the inevitable variability of neurostimulation parameters over time. Upon validation in other frameworks it increases the usability of this technology through decreasing the time and the cost of the treatment supporting the potential for future widespread use. This work suggests the exploitation of AI-based methods for developing the next generation of ‘smart’ neuromodulation devices.