Abstract
AbstractDespite over a decade of promising results, closed-loop neurostimulation for the treatment of drug-resistant epilepsy (DRE) still relies on manual parameter tuning and yields unpredictably variable outcomes. Fully automated algorithms with predictable outcomes have remained a theoretical possibility, mainly due to a lack of generalizable models of the brain’s dynamic network response to stimulation under varying parameters. In this work, we study predictive dynamical models of human intracranial EEG (iEEG) response under parametrically-rich neurostimulation and develop models that are predictively accurate and biologically interpretable. Using data fromn= 10 subjects, we show that evoked iEEG dynamics are best explained by stimulation-triggered switched-linear models with approximately 300ms of causal historical dependence. These models are highly consistent across stimulation amplitudes and frequencies, including STIM OFF durations, which allows for learning a single generalizable model from abundant STIM OFF and limited STIM ON data. Despite significant heterogeneity among subjects, we observed a consistent distance-dependent pattern whereby stimulation impacts the actuation site and nearby regions (≲ 20mm) directly with little or no network mediation, reaches medium-distance regions (20 ∼100mm) indirectly through network interactions and hardly reaches more distal areas (≳ 100mm). Peak involvement of network interactions occurs at about 60-80mm from the stimulation site. Due to their predictive accuracy and mechanistic interpretability, these models have far-reaching applications in model-based seizure forecasting and stimulation design for closed-loop neurostimulation.
Publisher
Cold Spring Harbor Laboratory
Cited by
1 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献