Bayesian adaptive design for neural recruitment curves

Abstract

Nonlinear response models remain ubiquitous across disciplines with recruitment curves describing muscle response to neural stimulation. Conventional acquisition methods typically rely on static uniform sampling, requiring many stimuli, while adaptive threshold-hunting designs such as ML-PEST forfeit accurate characterization of the overall curve. We present a Bayesian adaptive design that uses posterior uncertainty over curve parameters to propose stimuli that maximize information gain across the full recruitment curve. We additionally implement a focused acquisition method that can target arbitrary subset of curve parameters, which we use to evaluate and interpret our design. In simulations, our design outperforms uniform sampling in recovering threshold while matching ML-PEST, and outperforms ML-PEST in predictive curve accuracy while matching uniform. It likewise matches our threshold-focused acquisition, suggesting our design doesn't trade off threshold estimation for recovering the curve. We further show data collected using our design yields greater power when detecting a change in threshold in intervention studies. Our design proposes the next stimulus within 5 seconds, making it feasible for transcranial magnetic stimulation (TMS) experiments. Finally, we validate our design in TMS experiments and provide a library for Python, called onMEP, for online acquisition of neural recruitment curves.