Failure modes and mitigations for Bayesian optimization of neuromodulation parameters.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Biomedical Engineering

Background:

• Precision medicine aims to personalize neuromodulation treatments.
• Optimizing stimulation parameters is crucial but challenging due to noisy neuro-psychiatric data.
• Existing Bayesian optimization methods are not robust for high-noise, low-effect-size applications.

Purpose of the Study:

• To assess the sufficiency of standard Bayesian optimization for neuromodulation.
• To develop robust optimization techniques for precise neuromodulation parameter selection.
• To address the challenges of noise and safety in optimizing neuro-psychiatric outcomes.

Main Methods:

• Literature review of individual neurological and psychiatric effects.
• Development of simulated patient responses with varying signal-to-noise ratios.
• Assessment of standard Bayesian optimization and modified boundary-avoiding kernels.

Main Results:

• Standard Bayesian optimization failed for effect sizes below Cohen's d = 0.3.
• Over-sampling of parameter space boundaries contributed to optimization failures.
• A modified approach using input warp and boundary-avoiding kernels achieved robust optimization for effect sizes as low as Cohen's d = 0.1.

Conclusions:

• Standard Bayesian optimization requires caution in low-effect-size neuromodulation.
• Algorithms risk converging to local optima in noisy, low-signal environments.
• Boundary-avoiding techniques enhance robustness for personalized neuromodulation.