Computational models and tES dose optimization

Computational models of Tes—including tDCS, tACS, and elec-troconvulsive therapy (ECT), —can help to address two key issues for rigor and reproducibility, namely spatial targeting and individualization of dosing. Regarding spatial targeting of specific brain regions, tES is often rationalized based on modulating the activity of a specific brain region implicated in the illness, with the assumption that stimulating this brain region will bring about desired benefits. A majority of tES studies approach this challenge by placing a large (compared to the brain region) electrode on a scalp location broadly “over” the brain target. The second issue aided by computational models is the individualization of electrode placement. A majority of tDCS/tACS do not vary stimulation dose with the subject/patient, which may result in varied target modulation [63]. Electroconvulsive therapy (ECT) typically individualizes dosage by varying the duration and frequency of the stimulus train. However, the ECT pulse current amplitude and pulse width—key determinants of the induced stimulation strength in the brain—remain fixed across individuals. The fixed stimulus current amplitude results in differential dosing in the brain, potentially contributing to variability in outcome [64]. Without consistent modulation of clinical targets, the efficacy and reproducibility of tES trials may be suboptimal.

The strategies for addressing these limitations are both doctrinal and practical. The continued use of large electrodes placed on opposite sides of the head, which may result in current flow through extensive volumes of the cortex and deep brain [65,66] is encouraged by experience (e.g., positive outcomes from prior trials) and the simplicity of using two-electrode devices (e.g. sponges positioned with rubber straps for tDCS, or large steel disc electrodes for ECT [35]). Relatively few studies adopt High-Definition (HD) montages wherein arrays of smaller electrodes can steer current during tDCS [46,67–70] and tACS [71,72] for presumed increased focality [73]. Even with two large electrodes, there is significant sophistication in the use and optimization of approaches using two large electrodes either to intentionally engage a broad network [74,75] or maximally stimulate a given brain region without necessarily optimized focality [76–80]. Nonetheless, computational models are important to rationalize and quantify the stated hypothesis of a tES trial. Given this ubiquitous need, access to robust and simple-to-use modeling software, including software that can automatically process imaging data in a manner that is suited for current flow modeling, represents a gap, in contrast to the ready availability of conventional image segmentation tools.

Over a decade, significant progress has been made in translating computational models to practice [81,82]. With regard to model validation, numerous studies [73,83–85] have confirmed the general model predictions illustrated in Fig. 1–that large electrodes produce diffuse current flow, while small electrode arrays may yield categorical increases in focality. Notably, intracranial recordings in humans demonstrate that models are fairly accurate in predicting distribution of electric fields across the brain (with correlation of predicted and measured fields around r = 0.81) [86]. Neurophysiological studies have also confirmed that individual differences can be predicted and controlled through the use of models [73]. In pediatric studies, computational models have suggested a need for reduced stimulation intensity [38,87]. Computational models have been used to design montages to direct current flow through lesioned brains following stroke [60]. Ongoing efforts to increase access to computational models include basic graphical-user interfaces (GUI) [88], packaged engineering tools [89,90], the development of standards [91], and importantly, algorithms that will reduce the computational burden [78] and automate image processing for individual electric field modeling [92,93].

A. M1-SO configuration: Sponge electrodes, one over left primary motor cortex, one over the contralateral supraorbital ridge. B. Bilateral dorsolateral prefrontal cortex configuration: Sponge electrodes over the F3 and F4 EEG sites. C. 4 × 1 HD-tDCS M1 configuration: High-definition electrodes, one over M1, four return electrodes surrounding the center electrode. The electric field was simulated with a current amplitude of 1 mA. Electric field simulation was performed using SimNIBS 2.0.1 [191].

There are straightforward strategies for addressing remaining gaps in translating computational models into practice: educate the scientific community (e.g., journal and grant reviewers) regarding the role of computational models in hypothesis-driven tES research, support initiatives to create new tools, and promote the use of enhanced methodology. Failure to leverage computational models in tES research for pragmatic reasons can be addressed by providing and enhancing access to easy-to-use computational models that can design individualized and optimized montages for a given target region. Continued use of ad hoc electrode montages can be justified, for example, based on prior empirical success with a given montage, but claims that prior outcomes reflect modulation of a specific brain region may be hard to justify. “Functional targeting” [14] allows for modulation of an active network without targeted brain current flow, but the selection of stimulation dose should always be rationalized. Additional important innovations relate to computational neurostimulation, where models of current flow are linked to neuronal and, ultimately, behavioral models [94,95], and new algorithms link neurophysiological data with stimulation strategies (e.g., EEG-guided tES) [91,96–98]. Rising concerns about rigor and reproducibility render the adoption of computational models imperative, supporting consideration of when the use of conventional pad or HD montages are appropriate. Uninformed and misguided electric fields are one of the many possible causes of variability in tDCS/tACS research [99–103] that can be readily constrained with the use of computational models. Importantly, recognizing that computational models are an evolving tool to support rational hypothesis-driven experimentation (not ends in themselves) makes these models pivotal in enhancing the rigor and reproducibility of tES research.

Important questions remain about the utility of models either for montage design for a trial or individualizing current per subject [104]; however, these unknowns are not an excuse to not use models to the extent practical. For example, an important challenge is relating regional brain current flow with resulting changes in neuronal information processing and ultimately behavior. Efforts to bridge dose to behavior, also called computational neurostimulation, are ongoing. At the moment, the (implicit) assumption across applications using models is that brain regions respond in a monotonic/linear fashion with local current flow (electric field) intensity [105], such that increasing current delivered to a given brain region increases efficacy regardless of brain state and disregarding connectivity with other brain regions. Although this assumption is increasingly challenged by dose-response studies [106,107], at a more basic level one can assume brain regions receiving little current flow are spared direct effects of stimulation. For all these open questions on how to leverage models, they remain readily accessible and useful tools to support hypothesis-driven trials and indeed address questions on dose-response.