Group optimization methods for dose planning in tES

Abstract

Objective: Optimizing transcranial electrical stimulation (tES) parameters-including stimulator settings and electrode placements-using magnetic resonance imaging-derived head models is essential for achieving precise electric field (E-field) distributions, enhancing therapeutic efficacy, and reducing inter-individual variability. However, the dependence on individually personalized MRI-based models limits their scalability in some clinical and research contexts. To overcome this limitation, we propose a novel group-level optimization framework employing multiple representative head models.

Approach: The proposed optimization approach utilizes computational modeling based on multiple representative head models selected to minimize group-level error compared to baseline (no stimulation). This method effectively balances focal stimulation intensity within targeted brain regions while minimizing off-target effects. We evaluated our method through computational modeling and leave-one-out cross-validation using data from 54 subjects and analyzed the effectiveness, generalizability, and predictive utility of anatomical characteristics.

Main results: Our approach demonstrated that group optimization significantly outperformed protocols derived from standard templates or randomly selected individual models, notably reducing variability in outcomes across participants. Additionally, correlations between anatomical features (e.g., head perimeter and tissue volumes) and E-field parameters revealed predictive relationships. This insight enables further optimization improvements through the strategic selection of representative head models that are electro-anatomically similar to the target subjects.

Significance: The proposed group optimization framework provides a scalable and robust alternative to personalized approaches, substantially enhancing the feasibility and accessibility of model-driven tES protocols in diverse clinical and research environments.

Data access statement: The data that support the findings of this study are available from the corresponding author, R.S., upon reasonable request.

Keywords: computational models; dose parameter; transcranial electrical stimulation.

Figure 1:

The modeling steps performed in this study. (a) From left to right: Segmentation of the structural head MRIs into the different tissues, construction of a 3D head model for finite element calculation, mapping of the 10–10 EEG system electrode positions to the surface of the scalp, and calculation of the lead-field matrix for every bipolar combination of electrodes with Cz as a cathode. (b) The left dorsolateral prefrontal cortex (lDLPFC) target is defined in the head surface of the MNI template. The target comprises Brodmann areas 9 and 46. (c) The lDLPFC target in the grey-matter (GM) surface of each individual head model after mapping from the MNI space to the participant’s native space. (d) Distribution of the En-field in the cortical surface (GM surface) of each participant, induced by subject-specific protocols targeting the lDLPFC, was obtained by performing a personalized optimization of the normalized error with respect to no intervention (NERNI) for each participant. (e) Same as d, but now for the group-optimized protocol obtained from maximizing the average of NERNI across all participants. A common color scale was used for all the plots of En (in units of V/m). Scalp reconstructions have been anonymized.

Figure 2:

Different reference distances measured along the scalp of one of the participants. The different lines are determined by connecting reference anatomical points along the scalp of the participant: Nz (nasion), Iz (inion), R/LPA (right/left pre-auricular points).

Figure 3:

Average and standard deviation of the NERNI distributions obtained in the test participants when the protocols are obtained from different sources. The p-values were obtained with a post-hoc pairwise Dunn test with Bonferroni corrections. Only statistical significant differences (p-value<0.05) between the groups are shown.

Figure 4:

Distribution of the normalized error with respect to no intervention (NERNI) (a) and $<E_n>$ (b) induced by protocols derived from different approaches: personalized, group, Colin-based template, ICBM152-based template, and templates based on each one of the biophysical head models included in the analysis (non-pers_part_$<id>$).

Figure 5:

Distribution of the normal component of the E-field in the cortical surface of the participant where the group approach performed the best regarding the normalized error with respect to no intervention (left column) and for the participant where it performed the worst (right column). The figures show, in a common color scale (in units of V/m), from top to bottom: the personalized protocol, the group-optimized protocol (leave one out approach), the protocol obtained with the ICBM152 template, and the protocol obtained with the Colin template.

Figure 6:

Relationship between the normalized error with respect to no intervention (NERNI) and $<E_n>$ on the. left dorsolateral prefrontal cortex. The solid line indicates the fit to the data, and the red points indicate the data points obtained from the personalized approaches. The blue points indicate the other different optimization approaches: group (a), Colin-based template (b), ICBM152-based template (c), and templates based on each one of the biophysical head models included in the analysis (d).

Figure 7:

Linear regression of $<E_n>$ vs the difference in anatomical features between the template used to evaluate the protocol (non-personalized template, based on the biophysical head model of one of the participants) and the participant’s anatomical features. Features from left to right: difference between perimeter of the heads along an sagittal/coronal plane (sag/cor_perim_diff), difference between the volumes of the scalp (scalp_vol_diff), and difference between the volumes of the WM/GM normalized by dividing them by the sum of the volumes of all the tissues (WM/GM _vol_diff_norm).

Figure 8:

Linear regression of NERNI using as features all linear and second order terms resulting from combining differences in anatomical features between the template used to evaluate the protocol (non-personalized template, based on the biophysical head model of one of the participants) and the participant’s anatomical features. Only interactions resulting in $R^2$ values larger than 0.1 are shown. The relevant features include: ax/cor_perim_diff (scalp’s perimeter measured along the axial and coronal planes), ax/cor/sag_perim_diff (scalp’s perimeter measured along the axial, coronal and sagittal planes, and normalized to the sum of all 3 perimeters), scalp/wm_vol_diff (volume of the scalp/WM volume), and scalp/wm/gm/skull_vol_diff_norm (volume of the scalp/WM/GM/skull, normalized by the sum of the volumes of all the tissues).

Figure 9:

Plot of predicted and observed average $⟨E_n⟩$ and NERNI (top/bottom row, respectively) for every template, derived from the (other) head models of the participants. (a/c) Model fit with all anatomical features (perimeters and volumes); (b/d) Model fit with only the perimeters of the scalp.

Figure 10:

Extension of the group optimization to include repetitions of subjects with different tissue conductivity combinations based on known a priori distributions.