An augmented preference-based Bayesian approach for optimizing neuromodulation stimulation parameters using meta learning

<i>Background.</i>Electrical neuromodulation is increasingly used in the treatment of neurological disorders; however, the selection of stimulation parameters that provide optimal therapeutic benefits remains a major challenge. Moreover, identifying pathological biomarkers linking the effect of stimulation parameters to alleviating symptoms, and hence required for optimizing stimulation parameters, might not always be possible.<i>Objective.</i>We present an augmented, preference-based Bayesian optimization algorithm to optimize stimulation parameters for participants undergoing neuromodulation. This algorithm incorporates two key features: I) It prioritizes the participant's preferences for stimulation parameters, making it independent of the need for pathological biomarkers. II) It leverages meta learning, using historical participant data to guide the initial optimization for new participants and overcome initial data sparsity. This approach improves both prediction accuracy and convergence speed.<i>Approach.</i>Consider preference training data collected from a set of historical participants who share the same neurological disorder as a new (target) participant. Within that population, there may be different response phenotypes. The goal is to identify historical participants whose stimulation-response phenotype is most similar to the target participant, and leverage their data to accelerate and improve parameter optimization for the target participant. To achieve this, the algorithm iteratively performs a two-step process:(I) a novel, iterative weighting procedure that identifies historical participants with stimulation preferences closest to the target participant, and (II) meta learning that combines the training data of the identified participants with the limited training data of the target participant to train novel, augmented preference learning models. These models are then used to predict the stimulation parameters expected to maximize the target participant's preference.<i>Main</i><i>results.</i>The proposed algorithm has been validated using synthetically generated data sets that simulate participant preference behavior during neuromodulation.<i>Significance.</i>This approach holds promise for improving personalized neuromodulation therapies and advancing treatment outcomes for neurological disorders without the need for a tedious data collection process and disease-specific pathological biomarkers.