Personalized Transcranial Electrical Stimulation

• Personalized tES is a non-invasive neuromodulation approach that customizes stimulation parameters using individual anatomical and physiological data.
• It combines individualized forward modeling with inverse optimization to improve target engagement, focality, and dose consistency across diverse subjects.
• Physiology-informed adjustments and state-aware calibration refine parameter selection, addressing inter-individual variability in brain electric fields.

Personalized transcranial electrical stimulation (tES) denotes individualized non-invasive neuromodulation in which stimulation parameters such as montage, current distribution, waveform, frequency, phase, and timing are selected using subject-specific anatomy and, increasingly, physiology and brain state rather than fixed canonical protocols. The central motivation is that under identical device settings the induced brain electric field can differ markedly across people because of variability in scalp and skull thickness, cerebrospinal fluid distribution, cortical folding, lesions, developmental anatomy, conductivity, and ongoing neural dynamics. Personalized tES therefore combines individualized forward modeling with inverse optimization and, in some implementations, electrophysiological or behavioral calibration, with the aim of improving target engagement, focality, dose consistency, and safety across modalities including tDCS, tACS, HD-tES, and temporal interference stimulation (Wang et al., 1 Sep 2025, Mastakouri et al., 2017).

1. Conceptual basis and scope

The defining premise of personalized tES is that one-size-fits-all stimulation is often mismatched to the underlying biophysics and neurophysiology of the individual. Computational modeling work has framed this problem in terms of substantial inter-individual variability in head geometry, tissue conductivities, cortical topology, lesion or implant status, and ongoing brain state, all of which perturb the mapping from scalp current to intracranial field (Wang et al., 1 Sep 2025). A closely related physiological formulation appears in motor rehabilitation research, where subject-specific EEG–performance relationships varied in both strength and sign, especially in alpha, beta, and high gamma bands, directly arguing against a universal stimulation frequency or montage (Mastakouri et al., 2017).

This variability is not only anatomical. In pediatric HD-tDCS, standard montages produced robust age-dependent reductions in target electric-field intensity and global focality, and sex-dependent differences were mediated by tissue composition, with scalp volume consistently mediating sex effects and a joint multi-tissue mediation explaining age-related electric-field changes (Liu et al., 1 Dec 2025). This suggests that a nominally identical current dose can systematically over-dose one subgroup and under-dose another even before any physiological state dependence is considered.

Within this framework, personalization is broader than montage optimization alone. It includes anatomical personalization through individualized head models, physiological personalization through rhythm- or task-informed parameter selection, and, in closed-loop formulations, state-aware adaptation of stimulation in response to measured brain activity. The review literature explicitly characterizes this as a transition from empirical trial-and-error to dynamic and individualized stimulation planning (Wang et al., 1 Sep 2025).

2. Individualized forward models and dose prediction

The computational core of personalized tES is the forward problem. Under the quasi-static approximation used for conventional tES, the governing equations are

$$\nabla \cdot (\sigma \nabla \phi) = 0,\qquad \mathbf{E} = -\nabla \phi,\qquad \mathbf{J} = \sigma \mathbf{E},$$

with $\sigma$ representing tissue conductivity and with electrode boundary conditions supplied either as current fluxes, prescribed potentials, or a complete electrode model (Rashed et al., 2020). In practice, personalized pipelines construct a subject-specific conductor model from MRI, assign tissue conductivities, generate a finite-element or finite-difference model, and solve for the electric field in the target and non-target regions (Wang et al., 1 Sep 2025).

The fidelity of this forward model is highly sensitive to segmentation quality at specific tissue boundaries. A dedicated sensitivity analysis found that in brain, sensitivity to segmentation accuracy is relatively high in cerebrospinal fluid, moderate in gray matter, and low in white matter for both tES and TMS. A CSF segmentation accuracy reduction of $10\%$ in terms of Dice coefficient led to a decrease up to $4\%$ in normalized induced electric field, whereas a GM segmentation accuracy reduction of $5.6\%$ led to an increase of normalized induced electric field up to $6\%$; hotspot localization uncertainty was typically within a few millimeters for CSF and GM perturbations and zero for WM in the studied motor-cortex scenario (Rashed et al., 2020). These results localize uncertainty to the CSF–GM boundary and imply that sulcal fidelity is a first-order requirement for reliable dose planning.

Conductivity assignment is a second major uncertainty source. In vivo estimation from simultaneous intracerebral stimulation with scalp EEG and sEEG showed that estimated conductivities for the same patient changed with stimulation position, measurement position, and modality, explaining part of the heterogeneity in the literature (Altakroury et al., 2022). EEG measurements were more informative for scalp, skull, and CSF, whereas sEEG was more informative for GM and WM, and combined datasets improved complementarity but could be biased by sensor-count imbalance. For personalized tES, this means that even a geometrically accurate head model can remain miscalibrated if conductivities are treated as fixed atlas constants.

Several recent methods address the throughput bottleneck in individualized modeling. ForkNet generated multi-tissue segmentations from T1 MRI and SubForkNet extended this to seven deep brain structures, with high matching especially in Thalamus, Caudate, and Putamen and with small global electric-field errors when these segmentations were embedded into personalized tDCS head models (Rashed et al., 2020). CondNet went further by estimating voxel-wise, non-uniform conductivity maps directly from T1- and T2-weighted MRI, bypassing anatomical segmentation and yielding smoother electric-field distributions than uniform tissue-wise conductivity assignment (Rashed et al., 2019). Together, these approaches shift personalization from a labor-intensive manual pipeline toward a scalable imaging-to-conductor workflow.

Electrode modeling matters as well. Advanced boundary modeling compared the complete electrode model (CEM), the gap model (GAP), and the point electrode model (PEM) in a realistic auditory-cortex stimulation case and found that brain current densities were virtually identical across CEM/CEM, GAP/PEM, and CEM/PEM for realistic average contact impedance, whereas differences were concentrated in the skin compartment under electrodes (Agsten et al., 2016). The practical implication is specific: GAP is sufficiently accurate for brain-dose prediction and optimization, while CEM is required if skin current density, edge-current enhancement, or heating effects are part of the personalization objective.

3. Inverse optimization and individualized montage design

Once the forward operator is available, personalization becomes an inverse problem. The standard formulation exploits linear superposition, representing the field as a mapping from electrode currents to intracranial electric field, commonly written as $E = A I$ or, in directional formulations, as an ROI-weighted projection of $A I$ (Liu et al., 1 Dec 2025). Objectives typically trade off target intensity against focality, off-target suppression, depth, avoidance regions, and practical constraints such as current neutrality, per-electrode current bounds, total current budgets, and a limited number of active electrodes (Saturnino et al., 2019).

A general anatomical optimization result is that accessibility depends strongly on spatial position and constraints. Systematic cortical mapping showed that superficial gyral crowns are the most accessible, whereas sulcal, medial, and inferior targets exhibit larger targeting errors and broader spread; under $I_{\mathrm{ind}} = 1$ mA and $I_{\mathrm{tot}} = 2$ mA, maximal dose could reach about $\sigma$0 V/m at some gyral crowns but was typically $\sigma$1 V/m for sulcal and deeper targets, and practical focal single-target solutions required only about six to eight active electrodes (Saturnino et al., 2019). These results delimit what personalization can achieve within the physics of conventional TES rather than implying unrestricted steerability.

Optimization methodology has diversified. Pediatric HD-tDCS introduced a dual-objective framework in which a Genetic Algorithm seeded a Multi-Objective Particle Swarm Optimization to produce individualized Pareto fronts summarizing the trade-off between target electric-field intensity and focality (Liu et al., 1 Dec 2025). From these fronts, two prescriptions were defined: a dose-consistency strategy that enforced fixed target intensity across individuals and eliminated age and sex effects on target electric field, and a target-engagement strategy that maximized target intensity under safety limits and still showed no significant demographic effects on the achieved maximal electric field (Liu et al., 1 Dec 2025). This is a notable formalization of personalization as either variance reduction across subjects or target maximization within subjects.

Threshold-aware optimization has altered the focality objective itself. HingePlace extended traditional convex placement algorithms by using a symmetrized hinge loss that penalizes only off-target fields above physiologically meaningful tolerances rather than pushing all off-target fields toward zero. In simulations, HingePlace-designed montages generated more focal neural responses by as much as $\sigma$2 than traditional algorithms, with the reported gains arising from explicit use of neural thresholding behavior (Goswami et al., 3 Feb 2025). The conceptual shift is that personalization is not merely field shaping but neural-response shaping under a nonlinear activation model.

Multi-objective evolutionary optimization provides a complementary route when objectives are non-convex or simultaneous targeting is required. MOVEA generated Pareto fronts for HD-tACS and two-pair tTIS without manual weight adjustment, supporting joint optimization of intensity, focality, depth, avoidance zones, and steerability (Wang et al., 2022). In a different LP-based tradition, L1–L1 metaheuristic search optimized focality and intensity over a two-parameter lattice, with interior-point solvers producing smoother, more robust optima than simplex variants and with subject- and target-dependent sweet spots for the regularization parameters (Prieto et al., 2022). Across these frameworks, the common structure is individualized optimization under explicit safety and device constraints rather than reliance on fixed textbook montages.

4. Physiology-informed personalization and behavior-linked dosing

Anatomically individualized fields do not by themselves specify the most effective stimulation frequency, timing, or state dependence. Physiology-informed personalization attempts to close this gap by using measured neural activity or behavior to infer subject-specific stimulation parameters. A clear example fused brain–computer interface training with TES by learning subject-specific linear EEG decoders that predicted trial-by-trial movement smoothness during a 3D reaching task and then using the learned spatial and spectral weights to identify candidate stimulation sites and rhythms (Mastakouri et al., 2017).

In that framework, motor performance was quantified by normalized averaged rectified jerk,

$\sigma$3

and individualized decoders were learned by transfer learning: a prior model trained on 25 subjects was updated with the first 20 trials of the held-out subject (Mastakouri et al., 2017). The personalized model significantly predicted subjects’ final mean movement smoothness with $\sigma$4, $\sigma$5, whereas the prior-only model did not ($\sigma$6, $\sigma$7), and trial-wise coherence was marginally significant for the updated model but non-significant for the prior model (Mastakouri et al., 2017). The key relevance for personalized tES is that rhythm–performance relationships varied in both topography and sign across subjects, especially in alpha, beta, and high gamma, so a fixed contralateral M1 protocol could average away genuine but opposing individual effects.

The same study proposed a concrete decision rule for individualized tACS: select bands and sites where increased power is associated with smoother movements, prioritize features that are relevant in both decoding and encoding analyses, and then experimentally test the resulting candidates because the models remain correlational (Mastakouri et al., 2017). This causal caution is central. A decoder can restrict the search space, but it does not prove that increasing the implicated rhythm at the implicated site will improve behavior.

At a more theoretical level, frequency personalization can also be cast as a network-dynamics problem. A cortical network model of interrupted depolarizing stimulation found optimal desynchronization near $\sigma$8 Hz, with lower frequencies increasing synchrony and very high frequencies becoming ineffective at fixed amplitude (Schütt et al., 2012). The model was superthreshold and therefore does not directly translate to human subthreshold tES, but it provides a mechanistic rationale for selecting moderately high frequencies when the goal is to fragment synchronized recruitment rather than entrain ongoing rhythms. A plausible implication is that physiology-informed personalization must distinguish between protocols intended to reinforce endogenous oscillations and protocols intended to disrupt them.

5. Deep-target personalization and temporal interference stimulation

Classical TES has a depth–focality trade-off that severely constrains selective deep targeting. Systematic accessibility analysis concluded that it is not feasible to target deep brain areas with TES in general, although focusing the field in some specific deeper locations might be possible due to favorable conductive properties in the surrounding tissue (Saturnino et al., 2019). Temporal interference stimulation was developed against this background as a way to steer a low-frequency envelope into deeper regions while using kHz carriers at the scalp (Vassiliadis et al., 16 Dec 2025).

The tTIS principle is based on superposition of two high-frequency currents,

$\sigma$9

with a low-frequency envelope at $10\%$0. A commonly used envelope-amplitude metric is

$10\%$1

which is maximized by balancing magnitudes and co-aligning vectors within the region of interest (Vassiliadis et al., 16 Dec 2025). Personalized tTIS therefore depends heavily on individualized head models, finite-element analysis, subject-specific electrode positions, channel-current ratios, and target-specific choice of $10\%$2 for theta-, beta-, gamma-, or DBS-like envelopes (Vassiliadis et al., 16 Dec 2025).

Human evidence has shifted tTIS from a purely computational proposal to an early translational modality. Reported findings include hippocampal steering across subregions by altering channel current ratios, striatal modulation of BOLD and motor learning, and clinical pilot work in Parkinson’s disease in which 130 Hz tTIS targeting GPi reduced MDS-UPDRS III by about $10\%$3 points, or about $10\%$4, in a double-blind randomized study with $10\%$5 (Vassiliadis et al., 16 Dec 2025). More than 250 deep-target tTIS sessions were reported as well tolerated, with mild tingling and robust blinding against sham and active high-frequency controls (Vassiliadis et al., 16 Dec 2025). The field remains explicitly subthreshold in humans, but the translational emphasis has shifted toward state dependence, frequency selectivity, and individualized steering rather than brute-force intensity escalation.

A computational study of insula and hippocampus targeting makes the distinction between group-level and individualized deep-target optimization explicit. Using 60 high-resolution head models, it found that for the insula a group-level montage combining T7–P7 and Fp1–Fp2 achieved the highest focality and was comparable to individualized results with reduced variability, whereas for the hippocampus individualized optimization improved off-target suppression relative to the best group-level montage (Inoue et al., 27 Oct 2025). Stable group-level patterns required about 20 models for the insula and 9 for the hippocampus, and the current scaling needed to achieve 10% suprathreshold coverage at $10\%$6 V/m was far larger for the hippocampus than for the insula, quantifying the depth penalty (Inoue et al., 27 Oct 2025). The practical conclusion is target-dependent: group-level optimization may suffice for deep cortical targets such as the insula, but individualized tuning remains preferable for more deeply situated structures such as the hippocampus.

6. Validation, limitations, and unresolved questions

Despite technical progress, personalized tES is not a solved problem, and several recurrent misconceptions are contradicted by the current literature. The first is that personalization can be reduced to anatomical targeting alone. In fact, the field increasingly treats anatomy, conductivity, neural orientation, endogenous rhythms, and task or brain state as jointly relevant variables, and closed-loop or multimodal designs are being advanced precisely because field strength alone does not determine the behavioral effect (Wang et al., 1 Sep 2025). The second is that individualized field modeling automatically guarantees causal efficacy. Physiological decoders and encoding maps remain correlational unless tested experimentally, a point made explicitly in EEG-guided motor rehabilitation work (Mastakouri et al., 2017).

Validation also remains partial. Forward models can be calibrated by intracranial or scalp measurements, but uncertainty persists in skull conductivity, tissue anisotropy, segmentation errors, electrode positioning, gel bridging, contact impedance, and frequency dependence at higher carrier frequencies (Wang et al., 1 Sep 2025). The literature on in vivo conductivity estimation shows modality- and configuration-dependent variability rather than a single definitive conductivity vector for a given person (Altakroury et al., 2022). This implies that uncertainty quantification and, in robust formulations, uncertainty-aware optimization are not optional add-ons but part of the modeling problem itself.

Clinical translation introduces additional layers. Pediatric HD-tDCS showed that sparse equivalents can be derived from dense optimized solutions without performance loss under the target-engagement strategy, which is important for deployment complexity, but the framework still assumes isotropic conductivities and static anatomy (Liu et al., 1 Dec 2025). Segmentation accelerators and conductivity networks reduce preprocessing burden, but they also introduce learned priors whose failure modes under scanner variation, pathology, or poor anatomical contrast require further characterization (Rashed et al., 2019, Rashed et al., 2020). Likewise, advanced boundary models indicate that brain-dose prediction and scalp safety prediction are not the same task: GAP can be sufficient for the former, whereas CEM is needed for the latter (Agsten et al., 2016).

Current research directions therefore converge on a more integrated notion of precision neuromodulation. The computational review literature emphasizes multimodal head modeling, network-informed inverse design, empirical calibration with intracranial or imaging data, adaptive or closed-loop control, cloud-scaled workflows, and digital twins that are updated as anatomy and physiology change (Wang et al., 1 Sep 2025). This suggests a mature definition of personalized tES: not merely individualized montage selection, but a model-based, uncertainty-aware, and increasingly state-dependent control problem whose success depends on the joint accuracy of forward physics, optimization, and biological target definition.