Statistical Model of Motor-Evoked Potentials

Abstract

Motor-evoked potentials (MEPs) are widely used for biomarkers and dose individualization in transcranial stimulation. The large variability of MEPs requires sophisticated methods of analysis to extract information fast and correctly. Development and testing of such methods relies on the availability for realistic models of MEP generation, which are presently lacking. This paper presents a statistical model that can simulate long sequences of individualized MEP amplitude data with properties matching experimental observations. The MEP model includes three sources of trial-to-trial variability: excitability fluctuations, variability in the neural and muscular pathways, and physiological and measurement noise. It also generates virtual human subject data from statistics of population variability. All parameters are extracted as statistical distributions from experimental data from the literature. The model exhibits previously described features, such as stimulus-intensity-dependent MEP amplitude distributions, including bimodal ones. The model can generate long sequences of test data for individual subjects with specified parameters or for subjects from a virtual population. The presented MEP model is the most detailed to date and can be used for the development and implementation of dosing and biomarker estimation algorithms for transcranial stimulation.

Fig. 1.

(A) MEP amplitude model structure. Statistical distribution of (B) inter-individual (p_1,i – p_5,i) and (C) intra-individual (ε_ymult,ij, ε_xadd,ij, ε_yadd,ij) parameters to represent, respectively, the spread of MEP features in the subject population and trial-to-trial variability.

Fig. 2.

Model-generated IO curves of two virtual subjects, including 500 uniformly distributed samples and corresponding average sigmoid curve (A and C), and 100,000 samples at a single stimulation strength point within the curve slope (B and D). These examples illustrate previously described MEP features such as approximately lognormal distribution around the threshold (B) and locally bimodal distribution around the threshold (D).