Figure 1: Bayesian adaptive acquisition of recruitment curves using posterior uncertainty over curve parameters to maximize information gain.(a) Example MEPs from a real participant’s APB muscle in TMS experiment, recorded at uniform intensities (left), and at intensities proposed by the expected information gain (EIG) design (right). Both uniform and EIG are block randomized and start with the same initial samples (black dots)—eight samples at zero intensity (not counted) and three uniform samples. The abscissa is specified as percent maximum stimulator output (% MSO) for TMS but may instead use units of current, such as μ\\muA or mA, for spinal cord stimulation. Inset: schematic quantification of MEPs using either peak-to-peak (pk-pk) amplitude or area under the curve (AUC). (b) Recruitment curves estimated from uniform and EIG samples at the end of the experiment, with 26 samples each. Samples are colored by acquisition order from earlier (dark) to later (light). EIG yields a narrower threshold posterior (bottom), indicating greater certainty in threshold estimate, and allocates samples to more informative regions of the curve. Side panel: posterior distribution of saturation. (c) Starting with three initial samples, the EIG design updates the posterior over curve parameters and uses it to propose the intensity for the next sample (cross mark). It samples near the threshold early and later samples the transition regions to estimate the slope and saturation. Accordingly, the threshold posterior (bottom) narrows early, while the saturation posterior (right) narrows later. Neural recruitment curves characterize nonlinear muscle response to neural stimulation. They are derived using electromagnetic and electrical stimulation techniques, such as transcranial magnetic stimulation (TMS), spinal cord stimulation (SCS), and nerve stimulation, which are used to assess and modulate the neural systems that control movement ([Stefan et al., 2000](https://arxiv.org/html/#bib.bib29); [Chen et al., 2008](https://arxiv.org/html/#bib.bib27); [Picht et al., 2011](https://arxiv.org/html/#bib.bib32); [Millet et al., 2011](https://arxiv.org/html/#bib.bib26); [Bunday and Perez, 2012](https://arxiv.org/html/#bib.bib30); [MacDonald et al., 2013](https://arxiv.org/html/#bib.bib31); [Sayenko et al., 2015](https://arxiv.org/html/#bib.bib33); [Hofstoetter et al., 2021](https://arxiv.org/html/#bib.bib34); [Greiner et al., 2021](https://arxiv.org/html/#bib.bib25); [Pal et al., 2022](https://arxiv.org/html/#bib.bib24); [Arora et al., 2022](https://arxiv.org/html/#bib.bib36); [McIntosh et al., 2023](https://arxiv.org/html/#bib.bib11); [Balbinot et al., 2023](https://arxiv.org/html/#bib.bib35); [Koponen et al., 2024](https://arxiv.org/html/#bib.bib28)). When targeted to the sensorimotor system, these techniques produce motor-evoked potentials (MEPs) (Fig. [1](https://arxiv.org/html/#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Bayesian adaptive design for neural recruitment curves")a), which are quantified by their size, typically in terms of peak-to-peak amplitude or area under the curve (Fig. [1](https://arxiv.org/html/#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Bayesian adaptive design for neural recruitment curves")b, gray and blue dots). The recruitment curve (Fig. [1](https://arxiv.org/html/#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Bayesian adaptive design for neural recruitment curves")b, black curve) models the MEP size as a non-linear, non-decreasing function of stimulation intensity and exhibits a characteristic S-shape, with a steep increase above the threshold leading up to a plateau phase at high intensities (Fig. [1](https://arxiv.org/html/#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Bayesian adaptive design for neural recruitment curves")b, lower and side panels). Whereas parameters such as the slope and saturation may have specific neurophysiological interpretations, the threshold remains a critical parameter for inferring changes in corticospinal excitability ([Devanne et al., 1997](https://arxiv.org/html/#bib.bib12); [Farzan, 2014](https://arxiv.org/html/#bib.bib14); [Schicktanz et al., 2014](https://arxiv.org/html/#bib.bib41); [Vucic et al., 2023](https://arxiv.org/html/#bib.bib66); [Bestmann, 2024](https://arxiv.org/html/#bib.bib42)), while the overall recruitment curve characterizes the total corticospinal output over the stimulation range ([Carson et al., 2013](https://arxiv.org/html/#bib.bib49); [Potter-Baker et al., 2016](https://arxiv.org/html/#bib.bib43); [Kraus et al., 2018](https://arxiv.org/html/#bib.bib40); [Iyer and Madhavan, 2019](https://arxiv.org/html/#bib.bib45); [Kemlin et al., 2019](https://arxiv.org/html/#bib.bib46); [Filipović et al., 2021](https://arxiv.org/html/#bib.bib48); [Eginyan et al., 2023](https://arxiv.org/html/#bib.bib44); [Murphy et al., 2023](https://arxiv.org/html/#bib.bib47)). 出典: Bayesian adaptive design for neural recruitment curves(arXiv)
Figure 3: EIG proposes the next stimulus within 5 seconds, making it feasible for online TMS experiments.(a) Total computation time between consecutive stimuli for EIG and EIG focused on threshold when acquiring a single target muscle (left) and three muscles simultaneously (right). Boxes indicate the median time and span the interquartile range, whiskers span the (2.5, 97.5) percentiles, and dots indicate outliers. (b) Computation time for the posterior update step using the No-U-Turn sampler (NUTS), which accounts for most of the total computation time and gradually increases with the number of stimuli. (c) Computation time for the proposal step for EIG and EIG focused on threshold, which remains constant with the number of stimuli. EIG consists of two computational steps between consecutive stimuli—the update step, which estimates the posterior distribution over the curve parameters after collecting a new observation, and the proposal step, which computes the acquisition function over candidate stimuli to choose the next stimulus. The update step uses Markov chain Monte Carlo (MCMC) methods, such as the No-U-Turn sampler (NUTS) ([Hoffman and Gelman, 2011](https://arxiv.org/html/#bib.bib1)), whereas the acquisition function can be computed across candidate stimuli in parallel. To account for this, we implemented posterior updating on the CPU and acquisition evaluation on the GPU within our Python library onMEP. For multiple-muscle acquisition, the update and proposal steps are further parallelized across muscles, such that, with sufficient compute, the required time should approach that of a single muscle. 出典: Bayesian adaptive design for neural recruitment curves(arXiv)