Failure modes and mitigations for Bayesian optimization of neuromodulation parameters.

Precision medicine holds substantial promise for tailoring neuromodulation techniques to the symptomatology of individual patients. Precise selection of stimulation parameters for individual patients requires the development of robust optimization techniques. However, standard optimization approaches, like Bayesian optimization, have historically been assessed and developed for applications with far less noise than is typical in neuro-psychiatric outcome measures and with minimal focus on parameter safety.We conducted a literature review of individual effects in neurological and psychiatric applications to build a series of simulated patient responses of varying signal to noise ratio. Using these simulations, we assessed whether existing standards in Bayesian optimization are sufficient for robustly optimizing such effects.For effect sizes below a Cohen's d of 0.3, standard Bayesian optimization methods failed to consistently identify optimal parameters. This failure primarily results from over-sampling of the boundaries of the space as the number of samples increases, because the variance on the edges becomes disproportionately greater than in the remainder of parameter space. Using a combination of an input warp and a boundary avoiding Iterated Brownian-bridge kernel we demonstrated robust Bayesian optimization performance for problems with a Cohen's d effect size as low as 0.1.Our results demonstrate that caution should be taken when applying standard Bayesian optimization in neuromodulation applications with potentially low effect sizes, as standard algorithms are at high risk of converging to local rather than global optima. Mitigating techniques, like boundary avoidance, are effective and should be used to improve robustness.