Meta-Learning Boosts Neural Stimulation Model Robustness

Model-based closed-loop neural stimulation holds immense therapeutic potential, but its clinical deployment has been hampered by two main issues: models often fail catastrophically in predicting stimulation consequences, and per-session calibration is too time-consuming. This new research introduces a solution by applying meta-learning and pretraining to neural stimulation response modeling for the first time. The study extends Temporal Basis Function Models (TBFMs) with a novel architecture and algorithm based on Model-Agnostic Meta-Learning (MAML), evaluating it on optogenetic stimulation data from non-human primates. The results show a substantial reduction in catastrophic forecast failures; for instance, sessions with poor prediction accuracy dropped from 16 out of 40 to just 1 when using MAML-pretrained models. Prediction intervals also became significantly narrower. Crucially, the meta-learning approach reduced calibration requirements by 50-90% while maintaining accuracy. This efficiency gain makes previously infeasible experiments within clinical session-time constraints now possible. The findings provide empirical evidence that cross-session structure in stimulation responses is consistent enough to support pretraining, paving the way for more robust and sample-efficient closed-loop neural stimulation systems.