PRIME: A Neurophysiology-Informed Bayesian Optimization Framework for Adaptive TMS Motor Hotspot Mapping, Algorithm Design and Monte Carlo Evaluation

BackgroundAccurate primary motor cortex (M1) hotspot identification is a prerequisite for reliable transcranial magnetic stimulation (TMS) protocols, yet conventional grid search is stimulation-intensive and operator-dependent. Existing Bayesian optimization (BO) implementations, including BOOST and 3D-BOOST, do not model within-session neurophysiological variability. MethodsWe present PRIME (Probabilistic Response-guided Intelligent Motor Exploration), a closed-loop BO framework for three-parameter TMS hotspot mapping over coil position (X, Y) and orientation ({theta}), with five innovations targeting cortical excitability drift, amplitude-dependent noise, transient artifacts, sub-threshold MEP integration (10 {micro}V floor), and cross-subject GP prior warm-starting. These were evaluated alone and in combination with amplitude-weighted center-of-gravity (CoG) convergence and estimation across 18 configurations in a Monte Carlo simulation (30 subjects, 3 repetitions each). ResultsAlgorithm configuration significantly affected all outcomes (Friedman tests, all p < 0.001, Kendalls W = 0.50-0.65). MultiFid_CoG_10uV achieved the lowest median XY error (1.14 [0.62-2.12] mm; 50.9 {+/-} 6.2 stimuli; 96.7% convergence), a 68.2% error reduction and 41.5% stimulus reduction relative to grid Search (3.58 [2.66-5.12] mm; 87 stimuli). CoG estimation significantly reduced XY error relative to peak-response selection in 10 of 11 non-CoG configurations (largest gain: MultiFid, 58.7%; adjusted p < 0.001). ConclusionsNeurophysiology-informed BO improved simulated performance under the specified model, extending the 3-DOF approach of Grano et al. (2025) with explicit noise modeling and CoG-based convergence. MultiFid_CoG_10uV and DeltaBO_CoG are selected as candidates for prospective validation in a prospective triple-blind human study.