Anatomical misspecification in EMG Forward Models

EMG forward models increasingly generate synthetic data for machine learning, but the consequences of misspecifications in their inputs remain poorly understood. We systematically quantified the effects of anatomical and procedural misspecification in a state-of-the-art EMG simulator using manually segmented forearm MRI from 11 subjects. We perturbed anatomy, MRI resolution, fibre-generation protocol, and electrode placement, evaluating effects on motor unit action potential (MUAP) libraries and a downstream deep learning decomposition task. Subject-specific anatomy was the dominant source of variation: replacing a subject's forearm produced high library separability ($0.885$ vs. $0.448$ matched floor) and reduced decomposition yield by $7.8$ $\pm$ $1.1$ motor units. A cohort-mean anatomy performed no better than another individual's. Conversely, some perturbations strongly altered MUAP libraries without affecting decomposition. Pose-specific anatomy was also not recoverable from cohort-average corrections. These results demonstrate that simulator misspecification is task-dependent, but subject-specific anatomy remains a major determinant of synthetic EMG fidelity.