Neuromodulation for post-stroke aphasia rehabilitation: established cortical stimulation and emerging network-based targets

1 Introduction

Aphasia is an acquired impairment of language that most often follows damage to the language-dominant hemisphere after a stroke. It may affect speaking, comprehension, reading, writing, and pragmatic language use, with consequences extending far beyond impairment-level test scores. Approximately one third of people with acute stroke meet criteria for aphasia, although estimates depend on case definition and sampling (Engelter et al., 2006). Persistent aphasia limits autonomy, social participation, return to work, and access to health care, and is associated with reduced quality of life. Recovery is typically fastest during the first weeks and months, but meaningful treatment-related change remains possible years after stroke.

Speech-language therapy (SLT) is the foundation of aphasia rehabilitation. Systematic and randomised evidence supports structured, sufficiently intensive language intervention, including in chronic aphasia (Bhogal et al., 2003; Brady et al., 2016, 2025; Breitenstein et al., 2017). Nevertheless, the response is highly variable. Lesion location and volume, damage to connecting white-matter pathways, initial severity, time after stroke, domain-specific cognitive resources, treatment dose, and the fit between therapy and the impaired language process all shape outcome (Brady et al., 2025). Access to sustained high-dose treatment is also unequal (Brady et al., 2025). These realities have motivated interest in interventions that might increase experience-dependent plasticity during SLT (Davidson et al., 2024; Dawson et al., 2024).

Non-invasive neuromodulation alone does not restore language. Recent reviews have emphasised that non-invasive brain stimulation should be considered an adjunctive intervention designed to enhance the effects of evidence-based speech-language therapy rather than a stand-alone treatment (Harvey and Hamilton, 2022). Its clinically plausible role is to alter the physiological conditions under which a person practices language, thereby changing the probability, magnitude, or durability of learning. Repetitive transcranial magnetic stimulation (rTMS) and transcranial direct current stimulation (tDCS) have accumulated the largest evidence base (Lefaucheur et al., 2017; Elsner et al., 2019; Marangolo, 2020, 2025; Raymer and Johnson, 2024; Tan et al., 2024). More recent work has extended stimulation beyond conventional perisylvian targets to the cerebellum, primary motor and prefrontal cortices and spinal sensorimotor pathways (Marangolo, 2020, 2025). Transcranial photobiomodulation (tPBM) has additionally been proposed to influence cellular metabolism and neurovascular function (Naeser et al., 2020; Estrada-Rojas and Cedeño Ortiz, 2023).

Enthusiasm should be balanced against recurring limitations. Aphasia trials are often small, use heterogeneous interventions and outcomes, and preferentially report trained naming. A statistically significant improvement on practised items does not necessarily generalise to untrained words, connected speech, everyday communication, participation, or quality of life. Sham methods and masking also differ across devices, while crossover designs may be vulnerable to carryover when washout periods are short. In early post-stroke studies, spontaneous recovery and concurrent rehabilitation complicate causal inference. Finally, simple rules such as “excite the left hemisphere and inhibit the right” are increasingly difficult to reconcile with patient-specific network data.

This review critically updates the evidence through July 2026. We examine established cortical approaches and emerging network-based targets, distinguish therapeutic trials from mechanistic demonstrations, and organise the literature around three questions: what has been shown, what remains uncertain, and what evidence would be required for clinically responsible personalisation.

2 Methods of evidence identification and synthesis

We performed a structured narrative review of human research on neuromodulation for adult post-stroke aphasia. PubMed/MEDLINE was searched from database inception through 18 July 2026 using combinations of terms for stroke, aphasia, language rehabilitation, rTMS, theta-burst stimulation (TBS), tDCS, transcranial alternating current stimulation (tACS), cerebellar stimulation, transcutaneous spinal direct current stimulation (tsDCS), and photobiomodulation. Reference lists of relevant reviews and the full-text articles supplied with the manuscript were also screened. Priority was given to randomised or sham-controlled intervention studies, systematic reviews, meta-analyses, safety guidance, and mechanistic studies that directly informed target selection. Studies of primary progressive aphasia were excluded because the underlying neurodegenerative pathology and progressive course differ fundamentally from post-stroke aphasia, leading to distinct mechanisms of language impairment and potentially different responses to neuromodulation; mixed-aetiology samples were retained only when the post-stroke subgroup was identifiable or when the study illustrated a methodological issue. Protocols and non-peer-reviewed preprints were not used to support efficacy claims.

As this manuscript is a narrative review, the literature search and study selection were not conducted according to a formal systematic-review framework such as PRISMA. Accordingly, we did not perform a formal risk-of-bias assessment using standardized tools such as Cochrane RoB 2 or CASP. Instead, studies were critically appraised with respect to study design, sample characteristics, stimulation parameters, control conditions, outcome measures, and major methodological limitations. This approach was chosen to allow an integrative and critical discussion of the heterogeneous neuromodulation approaches and study designs represented in the aphasia literature. Because the objective of the review was interpretive rather than quantitative, we did not conduct a new meta-analysis.

3 A network and learning framework for aphasia neuromodulation

Language recovery after stroke reflects reconfiguration within a distributed, bilateral system rather than the behaviour of a single “language centre.” Indeed, longitudinal imaging has described an early reduction in language-network activity, followed by transient recruitment of bilateral or right-hemisphere regions and, in many patients, later re-engagement of residual left-hemisphere tissue (Saur, 2006; Hartwigsen and Saur, 2019). This group-level sequence is not a universal prescription. The contribution of the right hemisphere depends on which left-hemisphere regions are damaged rather than on lesion volume, premorbid lateralisation, time since stroke, task demands, and the remaining structural routes (Turkeltaub et al., 2026). Contemporary patient-level work demonstrates that right-hemisphere language-network recruitment can support performance in some people with aphasia rather than merely interfere with the damaged hemisphere (Raboyeau et al., 2008; Turkeltaub et al., 2026).

The influential interhemispheric-competition model proposed that loss of left-hemisphere output releases right-hemisphere homologues from transcallosal inhibition, after which excessive right-to-left inhibition impedes recovery (Rosen et al., 2000; Kiran, 2012). This model motivated inhibitory 1-Hz rTMS or cathodal tDCS over the right inferior frontal gyrus and excitatory stimulation of residual left-language cortex (Lefaucheur et al., 2017, 2020; Raymer and Johnson, 2024). Those strategies remain biologically plausible in selected patients, but the model should be regarded as one possible network state, not a default explanation for all aphasia. Suppressing the right hemisphere may be unhelpful when it is carrying compensatory language or domain-general control functions. Conversely, excitation of the structurally disconnected left cortex may not always engage the intended network.

The physiological labels “excitatory” and “inhibitory” also require caution. High-frequency rTMS and intermittent TBS often increase corticospinal excitability in the healthy motor cortex, whereas low-frequency rTMS and continuous TBS often reduce it (Tan et al., 2024). Similarly, anodal tDCS tends to depolarise and cathodal tDCS to hyperpolarise neuronal membranes in simplified preparations (Lefaucheur et al., 2017; Elsner et al., 2019; Marangolo, 2020, 2025). In the lesioned brain, however, response depends on orientation of the induced field, ongoing activity, local inhibitory-excitatory balance, medication, dose, prior stimulation, and network connectivity (Nitsche et al., 2003; Poreisz et al., 2007; Monte-Silva et al., 2013; Bikson et al., 2016). tDCS modulates the probability of firing rather than directly eliciting action potentials, and both electrodes contribute to the electric field (Nitsche and Paulus, 2000). Lesions and cerebrospinal fluid spaces can substantially distort current flow (Galletta et al., 2015). Polarity and frequency should therefore describe protocols, not predetermine their behavioural effects.

A learning framework offers a more clinically useful interpretation. Neuromodulation changes the state of a network while SLT specifies what is learned. The same stimulation may facilitate, impair, or leave behaviour unchanged depending on the task performed during the stimulation window (Marangolo, 2020; Shah-Basak et al., 2023). Pairing should be mechanistically coherent: phonological treatment should engage phonological computations, syntactic treatment should repeatedly recruit the relevant sentence network, and motor-speech practice should activate speech-motor circuitry (Marangolo, 2025). Dose must likewise be considered jointly; stimulation session number, treatment intensity, item exposure, and rest intervals are inseparable components of the intervention (Fregni et al., 2021; Marangolo, 2025). Table 1 summarises the principal neuromodulation approaches with respect to their primary physiological action, typical aphasia protocols, the current clinical evidence supporting their use, and the main unresolved questions.

Table 1. Neuromodulation modalities considered in post-stroke aphasia rehabilitation.

SLT, speech-language therapy; rTMS, repetitive transcranial magnetic stimulation; TBS, theta-burst stimulation; tDCS, transcranial direct current stimulation; tsDCS, transcutaneous spinal direct current stimulation; tACS, transcranial alternating current stimulation; tPBM, transcranial photobiomodulation.

4 Transcranial magnetic stimulation

4.1 Mechanisms and canonical protocols

TMS uses a rapidly changing magnetic field to induce an electric field in the superficial cortex. A single pulse can transiently perturb or probe a cortical region; repeated pulses can alter excitability and network coupling beyond the stimulation period. Most aphasia studies have used 1-Hz rTMS over the right pars triangularis to reduce activity in a contralesional inferior-frontal target. Other studies have applied higher-frequency rTMS or intermittent TBS to residual left-hemisphere language cortex, continuous TBS to right-hemisphere cortex or targets selected from functional imaging (Tan et al., 2024).

The relative focality of TMS is an advantage for hypothesis-driven targeting, and neuronavigation can improve anatomical reproducibility. Disadvantages include cost, acoustic and somatosensory sensations that complicate masking, operator dependence, and a small seizure risk (Lefaucheur et al., 2020) TBS delivers patterned pulses in a shorter session and is attractive for rehabilitation scheduling, but short administration time does not remove the need for careful dose reporting or safety screening (Tan et al., 2024).

4.2 Therapeutic evidence

Early reports of right inferior-frontal 1-Hz rTMS described naming improvements in small chronic-aphasia cohorts, sometimes persisting for months (Naeser et al., 2005a, 2005b, 2010; Martin et al., 2009). These studies were important demonstrations but frequently lacked sham control and reused participants across reports, limiting independence and protection against practice effects. Subsequently, small, controlled studies reported benefits on naming or global aphasia measures (Barwood et al., 2011; Weiduschat et al., 2011), but the samples were heterogeneous and typically comprised only 10–12 participants. A crossover pilot suggested changes in selected discourse word counts, although direct active–sham differences were limited and multiple outcomes were tested (Medina et al., 2012). Later trials produced a still more nuanced picture. Heikkinen et al. (2019) paired right-hemisphere 1-Hz rTMS with intensive language-action therapy in chronic aphasia. Language therapy improved performance, but active stimulation did not add a reliable benefit over sham. Georgiou and Kambanaros (2022) similarly found variable responses to right pars triangularis 1-Hz rTMS and continuous TBS in a six-person single-subject series. Ilkhani et al. (2018) reported pre-post improvement after low-frequency rTMS, but the absence of a sham comparator prevents attribution to stimulation.

Other protocols have targeted the residual left cortex. Kakuda and collaborators (2010a) found that 1-Hz rTMS (20 min, 10 sessions over 6 days, followed by weekly sessions for 3 months) administered to the left Wernicke’s area in two patients with fluent aphasia resulted in a slight improvement in several subtests of the Standard Language Test of Aphasia. Similar results were obtained in another study by the same authors (Kakuda et al., 2010b) in which rTMS was applied to the left frontal lobe in two of four poststroke nonfluent patients. In the open-label study of Szaflarski et al. (2011) functional-MRI-guided intermittent TBS over the residual left Broca’s area improved semantic fluency in six of the eight chronic participants enrolled (p = 0.028), although the absence of a sham condition prevents attribution to stimulation. Cotelli et al. (2011) reported improved object naming in three chronic non-fluent patients who received high-frequency rTMS over the left dorsolateral prefrontal cortex immediately followed by 25 min of individualised speech therapy over 4 weeks; because comparable gains also followed placebo rTMS plus speech therapy, the improvement cannot be attributed to stimulation. A single session of intermittent TBS over the left posterior temporal region produced a short-lived auditory-comprehension benefit in a 13-person pilot study, evident minutes rather than weeks after stimulation (Versace et al., 2020).

Imaging studies show that stimulation can alter network activity, but physiological changes are not, in themselves, evidence of rehabilitation efficacy. For example, Winhuisen et al. (2005) used PET and disruptive rTMS to test acute right-hemisphere language involvement, and Xu et al. (2021) used functional MRI after a single session of left primary motor cortex intermittent TBS. Both were mechanistic studies, not treatment trials.

The strongest recent evidence comes from controlled adjunctive designs and quantitative synthesis. In a four-arm study of 40 people with Broca-type aphasia, the combination of rTMS and SLT outperformed either component alone, although the modest sample and limited methodological reporting constrain certainty (Yaşa et al., 2023). A 2026 double-blind randomised trial used functional near-infrared spectroscopy to identify a left-hemisphere language target before neuronavigated 10-Hz rTMS; 15 sessions combined with intensive SLT produced a greater naming gain than sham plus SLT in subacute aphasia (Li et al., 2026). The study is encouraging because target selection was biologically informed and the stimulation was embedded in therapy, but replication and functional outcomes are needed.

A 2026 meta-analysis of 26 randomised trials (419 active and 318 control participants) reported an overall language benefit of rTMS, with the largest subgroup estimate for inhibitory stimulation of the right pars triangularis and for subacute participants (Wei et al., 2026). A separate meta-analysis of 12 randomised TBS trials involving 524 participants also found a pooled benefit but substantial heterogeneity (Zhang et al., 2026). Network meta-analysis of 26 trials and 1,136 participants ranked some multimodal or bilateral combinations highly, yet certainty was moderate to high mainly for standard low-frequency rTMS-plus-SLT comparisons and low or very low for several top-ranked novel combinations because estimates were indirect and imprecise (Lin et al., 2026). Rankings should therefore not be mistaken for definitive head-to-head superiority.

Overall, rTMS has the most mature evidence among non-invasive stimulation approaches for aphasia, but its effects are not uniform, and guideline-level endorsement for routine aphasia treatment remains limited (Lefaucheur et al., 2020). The best-supported clinical research model is repeated stimulation paired with structured SLT, using prespecified targets and outcomes. Table 2 distinguishes representative therapeutic trials from mechanistic studies that are often conflated in narrative summaries.

Table 2. Representative TMS studies and evidence classifications in post-stroke aphasia.

PWA: persons with aphasia; cTBS, continuous theta-burst stimulation; RMT, resting motor threshold; IFG, inferior frontal gyrus; iTBS, intermittent theta-burst stimulation; M1, primary motor cortex; PET, positron emission tomography; SLT, speech-language therapy; TMS–EEG, transcranial magnetic stimulation–electroencephalography; Level II: randomized controlled trial; Level IV: Case series/Single arm pre-post; N/A, not applicable.

5 Cortical transcranial direct current stimulation

5.1 Mechanisms, dosing, and targeting

tDCS applies a weak direct current between scalp electrodes. Conventional sponge-electrode montages are inexpensive, portable, and easy to combine with therapy, but their electric fields are diffuse. High-definition arrangements can improve model focus, although clinical superiority remains unproven. Typical aphasia protocols deliver 1–2 mA for approximately 20 min over 5–15 sessions while participants practice language. Left anodal, right cathodal, and bihemispheric montages have all been used; montage names alone are insufficient because electrode size, return-electrode location, lesion anatomy, and current direction determine the field (Lefaucheur et al., 2017, 2020; Esmaeilpour et al., 2018; Marangolo, 2025).

Unlike TMS, conventional tDCS does not directly evoke measurable neuronal responses but instead modulates ongoing neuronal excitability (Marangolo, 2020, 2025). Consequently, its effects are strongly influenced by the functional state of the stimulated network and are therefore typically evaluated in conjunction with behavioural therapy (Marangolo, 2020, 2025).

The interaction between neuromodulation and behavioural therapy is likely to depend not only on the stimulated cortical target but also on the specific neural processes engaged by the task performed during stimulation. Speech-language therapy provides a behaviourally induced state of network activation that may influence how TMS or tDCS modulates cortical excitability and activity-dependent plasticity. Rather than acting independently of ongoing neural processing, the effects of neuromodulation appear to depend, at least in part, on the functional state of the stimulated network and on the cognitive processes engaged during stimulation (Stagg and Nitsche, 2011; Fertonani and Miniussi, 2017). Thus, intensive and repetitive language practice may preferentially engage the residual neural circuits supporting the language trained function, providing a substrate on which neuromodulation can act.

This state-dependent interaction may be particularly relevant when the behavioural intervention is designed to repeatedly recruit a specific language or cognitive-linguistic process. For example, therapies targeting verb retrieval and thematic role processing, such as Verb Network Strengthening Treatment (VNeST), may repeatedly engage left inferior frontal and distributed language networks; importantly, studies combining VNeST with anodal tDCS over the left inferior frontal gyrus have provided preliminary evidence that stimulation delivered during this training may enhance treatment-related changes (Matar et al., 2022; Sheppard et al., 2025). In reading-based rehabilitation, iReadMore provides a particularly clear example of this principle: intensive reading practice engages distributed orthographic, phonological, semantic, frontal and occipital components of the reading network, while concurrent anodal tDCS over the left inferior frontal gyrus produced a small additional benefit on word-reading accuracy (Woodhead et al., 2018; Kerry et al., 2019). These examples suggest that the behavioural intervention should not be considered merely as an adjunct to stimulation, but as an important determinant of the neural state in which stimulation is delivered.

From this perspective, the therapeutic effect of neuromodulation may depend on the convergence between the stimulation-induced modulation, the residual structural and functional language network, and the language processes repeatedly engaged during therapy. Behavioural interventions that primarily target lexical-semantic retrieval, phonological processing, reading, action-verb processing, or sentence production may therefore provide different opportunities for state-dependent neuromodulation. Conversely, the same behavioural intervention may engage different residual networks across patients because of lesion location, network disconnection, and individual patterns of functional reorganisation. This interaction may contribute to the heterogeneity of neuromodulation outcomes observed across aphasia studies and supports the need to consider both the target of stimulation and the behavioural state induced during treatment when designing and interpreting combined neuromodulation–SLT interventions.

5.2 Evidence for language outcomes

Since word-finding difficulties are the most common symptom observed in aphasia, the majority of aphasia studies published through July 2026 have used a picture-naming task in combination with tDCS (Baker et al., 2010; Fiori et al., 2011; Flöel et al., 2011; Fridriksson et al., 2011, 2018; Kang et al., 2011; Marangolo et al., 2013a; Vestito et al., 2014; Shah-Basak et al., 2015; Wu et al., 2015; Norise et al., 2017; Silva et al., 2018; Rezaei et al., 2023; Alemanno et al., 2024; Bonilha et al., 2024; Ilardi et al., 2026). More specifically, there were 13 studies focused only on noun naming (Baker et al., 2010; Fiori et al., 2011; Flöel et al., 2011; Fridriksson et al., 2011, 2018; Kang et al., 2011; Shah-Basak et al., 2015; Wu et al., 2015; Norise et al., 2017; Silva et al., 2018; Rezaei et al., 2023; Alemanno et al., 2024; Bonilha et al., 2024) three studies on verb naming (Marangolo et al., 2013a; Fiori et al., 2019; Ilardi et al., 2026) and 2 in which both a noun and verb naming treatment approach was used (Fiori et al., 2013; Vestito et al., 2014) (see Table 3). Although picture naming has not been considered as an evidence-based treatment for aphasia, it has dominated the tDCS literature because it is readily standardised and repeatedly trained. It has been favoured because it promotes repetition and treatment intensity, thereby providing a constrained and replicable training method (Marangolo, 2025). Very few studies have combined tDCS with a theoretically driven treatment for verb production, such as the Verb Network Strengthening Treatment (VNeST) aimed to improve verb retrieval within different language contexts (Galletta and Vogel-Eyny, 2015), and ACTION, a cognitive protocol focused on verb inflexion and sentence construction (De Aguiar et al., 2015; Matar et al., 2022).

Table 3. Selected cortical tDCS studies informing the current evidence base.

PWA, persons with aphasia; BA6, Brodmann area 6; CETI, Communicative Effectiveness Index; IFG, inferior frontal gyrus; SLT, speech-language therapy; STG, superior temporal gyrus; TBI, traumatic brain injury; tDCS, transcranial direct current stimulation; VLSM, Voxel-Lesion-Symptom Mapping; Level II: randomized controlled trial; Level III-2: non-randomized controlled/comparative study with concurrent controls; Level IV: Case series/Single arm pre-post.

Independently of the electrode configurations (uni vs. bihemispheric), type of stimulation (left anodal tDCS vs. right cathodal tDCS) and stimulation intensity (1 mA vs. 2 mA) (see Table 3), all of these studies have pointed to an additional effect of tDCS on noun and verb naming within a multiple session paradigm (Baker et al., 2010; Fiori et al., 2011, 2019; Flöel et al., 2011; Fridriksson et al., 2011, 2018; Kang et al., 2011; Marangolo et al., 2013a; Vestito et al., 2014; Shah-Basak et al., 2015; Wu et al., 2015; Norise et al., 2017; Silva et al., 2018; Rezaei et al., 2023; Alemanno et al., 2024; Bonilha et al., 2024; Ilardi et al., 2026). Importantly, in some of these studies tDCS effects were still present at 1 to 4 weeks (Baker et al., 2010; Fiori et al., 2011, 2013; Flöel et al., 2011; Fridriksson et al., 2011; Marangolo et al., 2013a; Shah-Basak et al., 2015; Norise et al., 2017; Silva et al., 2018) up to 2 to 6 months (Vestito et al., 2014; Shah-Basak et al., 2015; Norise et al., 2017; Silva et al., 2018) post-intervention, thus, revealing a persistence of the language improvement due to tDCS over time. Beyond improvements in picture naming, Zhao et al. (2021) also reported that anodal tDCS combined with speech-language therapy produced greater language recovery than behavioural treatment alone in individuals with nonfluent post-stroke aphasia, further supporting the role of tDCS as an adjunctive intervention rather than a stand-alone treatment (see Table 3).

More recently, the phase-II randomised futility trial by Fridriksson et al. (2018) was designed to determine whether anodal tDCS over the preserved left temporal cortex, combined with computerised naming therapy, showed sufficient promise to justify further clinical evaluation. Although the active stimulation group improved by approximately six additional correctly named items compared with sham (adjusted between-group difference 5.7 words, 95% CI − 0.9 to 12.3), the confidence interval included the possibility of no effect. Consequently, the findings did not provide definitive evidence of efficacy but supported continued investigation in larger, adequately powered clinical trials. A later diagnostic-subgroup analysis suggested a greater benefit in participants with Broca aphasia, illustrating both the promise and the risks of post-hoc stratification (Bonilha et al., 2024).

Recent trials broaden the picture (Rezaei et al., 2023) compared four stimulation configurations in a small crossover study and found improved naming following individually selected inferior frontal or superior temporal stimulation. However, the findings should be interpreted cautiously given the limited sample size (n = 16). Alemanno et al. (2024) observed connectivity changes following right-anodal tDCS in 4 participants without a corresponding robust naming effect, underscoring that a neural biomarker cannot substitute for clinical benefit. Rubi-Fessen et al. (2024) reported larger gains during a period of individualised stimulation plus SLT than during an earlier period of SLT alone in early post-acute aphasia; because treatment order was fixed and recovery was ongoing, the effects of stimulation, cumulative therapy, and time could not be separated. Khairina et al. (2026) found mostly nonsignificant between-group differences in a quasi-experimental sample of 30 participants. Twardochleb et al. (2025) reported the greatest improvement in a combined tDCS-and-SLT group within a 90-person multicenter comparative study, but mixed stroke/traumatic etiologies, lack of sham, and non-equivalent treatment conditions limit aphasia-specific causal inference.

More rigorous and task-specific work is encouraging. Ilardi et al. (2026) randomised 32 people with chronic aphasia in a double-blind crossover study of 10 sessions of left inferior-frontal anodal versus sham tDCS during verb therapy. Active stimulation produced larger verb-naming gains that persisted for 1 month. Preserved anterior tissue and inferred tract integrity predicted response, supporting network-based stratification (Ilardi et al., 2026).

The application of tDCS in aphasia has also been extended beyond remediation of word-finding difficulties to the recovery of speech production (Marangolo et al., 2013b, 2014; Campana et al., 2015), articulatory difficulties (Marangolo et al., 2011, 2013c, 2016; Vila-Nova et al., 2019), sentence comprehension (Gilardone et al., 2026), word writing (Pisano et al., 2021) and reading (Woodhead et al., 2018; Cherney et al., 2021) (see Table 3).

Some studies have combined anodal tDCS over the LIFG with Conversational Therapy (Marangolo et al., 2013b, 2014; Campana et al., 2015). Similarly, Guillouët et al. (2020) demonstrated that combining tDCS with speech-language therapy significantly improved spontaneous speech production compared with sham stimulation, supporting the ecological value of neuromodulation beyond single-word naming tasks. Compared to the sham condition, the combined effects of tDCS with conversational therapy resulted in the production of more informative speech, which persisted at 1 month after the end of the treatment and extended to materials that had been administered at the beginning and end of the therapy sessions (Marangolo et al., 2013b, 2014; Campana et al., 2015). Similarly, in a study by Duncan and Nakkawita (2020), a small sample of four nonfluent PWA benefitted from a tDCS protocol combined with a storytelling task, which enhanced their ability to produce more informative content units compared to the sham condition (Duncan and Nakkawita, 2020). In a smaller double-blind crossover study, (Gilardone et al., 2026) paired left premotor anodal tDCS with 4 weeks of sentence-comprehension treatment in eight people with chronic nonfluent agrammatic aphasia. Both conditions improved, but active stimulation yielded larger model-adjusted comprehension gains and improvement on a functional communication measure. The four-month washout reduced, but did not eliminate, the interpretive limitations of a very small crossover sample. These findings are consistent with earlier evidence reported by You et al. (2011), who showed that cathodal tDCS applied over the right Wernicke’s homologue improved auditory comprehension in individuals with subacute post-stroke aphasia, suggesting that neuromodulation may facilitate language recovery beyond lexical retrieval.

Six studies demonstrated tDCS efficacy in the recovery of articulatory deficits when combined with nonword and word repetition (Marangolo et al., 2011, 2013c, 2016; Vila-Nova et al., 2019) or with Melodic Intonation Therapy (Vines et al., 2011; Cipollari et al., 2015; Yan et al., 2023) (see Table 3). Very recent studies have also demonstrated tDCS’s effectiveness in treating severe agraphia (Pisano et al., 2021) and central alexia (Woodhead et al., 2018; Cherney et al., 2021).

Meta-analytic estimates are positive but modest. A 2020 network meta-analysis concluded that some tDCS configurations may improve naming, while emphasising small trials, low certainty, and uncertainty across montages (Elsner et al., 2020). A 2026 meta-analysis of 12 randomised trials and 400 participants estimated a small naming effect overall (standardised mean difference 0.25, 95% confidence interval 0.05–0.45) and a larger estimate in chronic aphasia, but outcomes and protocols remained heterogeneous (Wang et al., 2026). Such pooled results support a signal, not a universal protocol.

Three conclusions follow. First, active tDCS should be tested as an addition to an evidence-based treatment for aphasia, with sham participants receiving the same therapy dose. Second, trained-item naming is a proximal endpoint; trials also need untrained items, discourse and functional communication outcomes, and persistency of the achieved gains. Third, apparent subgroup effects require prospective validation (Marangolo, 2020, 2025). Selected recent studies are summarised in Table 3.

6 Emerging network-based targets and modalities

6.1 Sensorimotor and domain-general cortical targets

Language production depends on sensorimotor and executive systems as well as classical perisylvian cortex (Gili et al., 2017; Hertrich et al., 2021; Shamma et al., 2021). Primary motor cortex can influence articulatory planning and action-language networks, while dorsolateral prefrontal cortex may support working memory, response selection, and cognitive control (Nagel et al., 2008; Tomaiuolo et al., 2021; Butler et al., 2023). These targets could be useful when therapy explicitly recruits their computations. However, evidence that modulation of a domain-general node improves everyday communication remains sparse, and improvements may reflect attention or response selection rather than the recovery of a language representation. Future studies should specify the predicted mediator and include non-language control measures to test whether effects are language specific.

6.2 Cerebellar stimulation

The posterolateral cerebellum participates in language prediction, timing, response selection, and error-based learning through crossed connections with cerebral association cortex (Macher et al., 2014; Mariën et al., 2014; Yuan et al., 2022). It is often structurally spared after a middle cerebral artery stroke, making it an accessible node from which to influence a damaged cerebral network. Polarity effects in the cerebellum are particularly difficult to infer, and protocols have used both anodal and cathodal stimulation (Grimaldi et al., 2014; Manto et al., 2022).

In a randomised crossover study of 12 people with chronic aphasia, right cerebellar cathodal tDCS enhanced verb generation but not verb naming, suggesting task specificity rather than a generalised lexical effect (Marangolo et al., 2018). Kim et al. (2024) used a double-blind within-subject design with 28 participants across anodal and cathodal cohorts and found improvement in qualitative dimensions of functional communication after cerebellar tDCS plus computerised treatment, but not in communication independence. Zheng et al. (2025) found no added Western Aphasia Battery language benefit in 22 actively treated participants with anodal tDCS of the right cerebellum paired with speech therapy compared with a nonconcurrent historical sham group, although some quality-of-life domains improved. The historical comparator weakens causal inference.

More recent evidence remains mixed. Xu et al. (2021) randomised 16 participants to cathodal cerebellar tDCS or sham with SLT. Unadjusted between-group differences were not significant; a baseline-adjusted analysis favoured active stimulation, and functional near-infrared spectroscopy showed connectivity changes. Carthery-Goulart et al. (2026) studied six Cantonese-speaking participants in a crossover design and found high variability without a consistent active-stimulation advantage. A secondary analysis of a 19-person crossover cohort found that integrity of cerebello-cortical white-matter pathways predicted naming response (Johnson et al., 2026). This biomarker result is valuable for mechanistic and stratification purposes, but does not, by itself, establish average efficacy.

Cerebellar tDCS is therefore a rational experimental adjunct, particularly for tasks involving response selection, sequencing, or error-based learning. Current samples are too small and outcomes too inconsistent for routine use. Trials should use concurrent randomised sham controls, lesion-aware tract measures, and therapy tasks chosen to engage the targeted cerebral-cerebellar circuit.

6.3 Transcutaneous spinal direct current stimulation

tsDCS applies direct current over the spinal cord and is proposed to influence ascending sensorimotor and cerebello-thalamo-cortical pathways. This “bottom-up” route is attractive because it avoids direct stimulation of the damaged cortex. In a 14-person crossover study, five sessions of thoracic anodal or cathodal tsDCS were compared with sham during language treatment; active stimulation selectively improved trained verb production at short follow-up (Marangolo et al., 2017). A subsequent functional connectivity study in 16 participants found anodal tDCS-related improvements in verb performance and changes in a cerebello-cortical network (Marangolo, 2020). Fourteen participants overlapped with the earlier cohort, so the reports are not independent replications. The verb specificity is mechanistically interesting, but efficacy remains unconfirmed in a parallel-group trial.

6.4 Transcranial alternating current stimulation

tACS delivers sinusoidal current to interact with ongoing neural oscillations. It offers a direct test of temporal coordination in language networks but raises difficult questions about individual frequency, phase alignment, current spread, and whether entrainment persists after stimulation (Grover et al., 2023). In a 13-person proof-of-concept study, 7-Hz high-definition tACS paired with language treatment yielded better production and stronger entrainment in an in-phase condition than sham (Keator et al., 2026). A separate 13-person study targeting theta-frequency activity reported a transient phonological short-term-memory benefit during stimulation, with marked responder heterogeneity and no durable group effect (Erickson et al., 2026). These studies support feasibility and mechanism; they do not establish rehabilitation efficacy. Replication should preregister frequency selection, masking assessment, phase condition, and delayed functional outcomes.

6.5 Transcranial photobiomodulation

Transcranial photobiomodulation (tPBM) delivers red or near-infrared light to the scalp. Proposed mechanisms include modulation of cytochrome-c oxidase, cellular energy metabolism, cerebral blood flow, oxidative stress, and inflammation. Most mechanistic evidence is indirect, and the dose reaching the human cortex varies with wavelength, power, scalp and skull properties, and device geometry.

Clinical aphasia evidence is preliminary. Naeser et al. (2020) reported on six people with chronic post-stroke aphasia treated across four sequential, nonrandomized light-emitting diode montage cohorts. Naming and functional connectivity changed in some configurations, but the design cannot separate treatment, practice, and montage effects. Estrada-Rojas and Cedeño Ortiz (2023) described one person who improved more during 5 months of tPBM plus SLT than during the preceding 5 months of SLT alone. The sequential single-case design is vulnerable to cumulative treatment, time, and expectancy effects. TPBM should currently be described as experimental; randomised, dosimetrically transparent, sham-controlled trials are required before claims of clinical benefit.

Table 4 summarises emerging approaches and their principal evidential constraints.

Table 4. Emerging network-based stimulation approaches in post-stroke aphasia.

PWA, persons with aphasia; IFG, inferior frontal gyrus; MTG/STG, middle temporal gyrus/superior temporal gyrus; LED, light-emitting diode; SLT, speech-language therapy; tACS, transcranial alternating current stimulation; tDCS, transcranial direct current stimulation; tPBM, transcranial photobiomodulation; tsDCS, transcutaneous spinal direct current stimulation. Level II: randomized controlled trial; Level IV: Case series/Single arm pre-post.

7 Safety, tolerability, and implementation

When delivered within contemporary guidance, non-invasive stimulation is generally well tolerated. Common tDCS and tACS effects include tingling, itching, warmth, transient erythema, headache, fatigue, or light flashes. Serious adverse events have not shown a causal signal within conventional low-intensity tDCS parameters, but reporting quality varies (Bikson et al., 2016; Antal et al., 2026). TMS commonly causes scalp discomfort, facial-muscle contraction, headache, and acoustic exposure. Seizure is rare but is the major acute risk, especially with higher-frequency protocols and in patients with predisposing factors (Rossi et al., 2021).

Stroke populations require more than a generic safety checklist. Screening should address seizure history, unstable medical conditions, cranial metal or implanted electronic devices, skull defects, medications that alter seizure threshold, pain and sensory impairment that may affect tolerability reporting, and cognitive or language limitations that affect consent. Hearing protection is required for TMS. Device-specific emergency procedures and trained supervision should be documented. Adverse events and withdrawals should be reported by randomised condition rather than described only as “well tolerated.”

Masking should also be treated as an outcome. Scalp sensations can differ between active and sham tES, and coil noise and muscle activation make TMS sham challenging. Trials should ask participants and therapists to guess allocation and report the success of blinding. For home-based stimulation, remote supervision may improve access, but only after reliable device locking, identity verification, skin checks, protocol adherence, emergency escalation, and equitable digital access have been demonstrated.

8 Factors contributing to heterogeneous outcomes

One of the most consistent observations in the reported literature is the marked variability in treatment effects across studies. This variability is unlikely to reflect a single methodological weakness. Rather, it arises from the interaction of biological, methodological, and analytical factors that influence both the neural effects of stimulation and their behavioural expression. Understanding these sources of heterogeneity is now more important than repeatedly asking whether neuromodulation “works,” because future progress will depend on identifying the conditions under which it is most likely to be effective.

8.1 Lesion-aware modelling and electric-field shaping

Biological variability probably represents the greatest challenge for neuromodulation research in aphasia. The label “post-stroke aphasia” encompasses patients with markedly different lesion topography, tissue viability, network disconnection, premorbid language organization, vascular burden, and recovery trajectories (Norise et al., 2017). Consequently, two individuals with comparable aphasia severity may exhibit profoundly different patterns of residual network organization and experience markedly different intracranial electric fields despite receiving the same stimulation protocol. These differences are likely to contribute substantially to treatment-response variability and cannot be captured by aphasia syndrome classification alone. Likewise, baseline severity should not be regarded solely as a prognostic factor but modelled appropriately, as it is intrinsically related to the potential for behavioural improvement (Harvey and Hamilton, 2022).

Importantly, anatomical variability can also alter the physical delivery of electrical stimulation. Stroke lesions introduce tissue compartments with electrical properties that differ from those of healthy brain tissue and may be accompanied by cortical atrophy, enlarged cerebrospinal-fluid (CSF) spaces, ventricular enlargement, and changes in tissue boundaries. Because current distribution depends on tissue conductivity and geometry, these alterations can modify the pathways through which current flows and consequently change both the magnitude and spatial distribution of the intracranial electric field (Bikson et al., 2012; Datta et al., 2012; Galletta et al., 2015; Saturnino et al., 2019). Indeed, computational modelling in a patient with post-stroke aphasia demonstrated that different electrode montages produced markedly different cortical electric-field distributions, highlighting the importance of considering the individual anatomy when interpreting the effects of tDCS (Galletta et al., 2015). In particular, CSF has substantially higher conductivity than brain tissue and can therefore act as a preferential pathway for current, potentially redistributing stimulation away from the intended cortical target. Stroke-related tissue loss and lesion location may consequently produce substantial inter-individual differences in electric-field strength and orientation even when identical scalp electrode montages and stimulation intensities are used.

This issue may be especially relevant when stimulation targets regions outside the canonical perisylvian language network, such as the cerebellum, supplementary motor areas, or other distributed nodes of the language network. In these cases, the relationship between electrode position and the neural tissue exposed to the strongest electric field may be less intuitive and more strongly influenced by individual anatomy. Computational approaches on finite-element modelling implemented in tools such as SimNIBS can incorporate individual anatomical information to estimate the electric-field distribution produced by different montages (Saturnino et al., 2019). Patient-specific modelling may therefore help determine whether a nominally identical montage actually produces comparable stimulation across patients and whether the intended neural target is effectively engaged.

These considerations suggest that personalization should not be restricted to selecting the cortical target according to lesion location or functional connectivity. It should also consider the individual electric field generated by the selected montage. Future studies should therefore combine lesion-aware head modelling with structural and functional connectivity measures to determine whether individualized electric-field optimization improves target engagement and, ultimately, clinical outcomes. Importantly, computational modelling should be treated as a means of improving the biological interpretability and reproducibility of stimulation rather than as evidence of clinical efficacy in itself. Prospective studies will be needed to establish whether model-guided electric-field shaping provides an incremental clinical benefit over conventional montage selection.

8.2 Intervention heterogeneity

A second source of variability arises from the interventions themselves. Across studies, stimulation protocols differ in target selection, stimulation parameters, treatment schedule, and timing relative to speech-language therapy. Equally important, behavioural interventions often differ substantially despite being described using similar terminology. For example, “anodal left IFG stimulation” may produce very different electric-field distributions depending on electrode placement and individual anatomy, while “naming therapy” may range from simple repeated picture naming to theoretically motivated lexical-semantic treatments. Consequently, apparent discrepancies between studies may reflect differences in the interventions being delivered rather than genuine inconsistencies in the biological effects of neuromodulation. Improving reproducibility will therefore require greater methodological harmonisation together with transparent reporting of both stimulation protocols and behavioural therapy.

8.3 Outcome and analytic heterogeneity

Outcome selection is another source of inconsistency. Studies vary widely in their choice of language measures, primary endpoints, follow-up intervals, and statistical approaches, making direct comparison difficult even when interventions are similar. Improvements on trained items, although sensitive to treatment effects, do not necessarily translate into generalised language recovery or functional communication. Future trials should therefore adopt clinically meaningful primary outcomes while systematically evaluating generalisation, participation, quality of life, and long-term maintenance. The development of core outcome sets would substantially improve comparability across studies.

8.4 Publication and sample-dependence problems

Finally, the published literature may overestimate treatment efficacy because small positive studies are more likely to be reported than negative ones. In addition, several influential publications derive from overlapping patient cohorts and therefore should not be interpreted as independent replications. Future systematic reviews should account for sample overlap, trial registration, and prespecified outcomes. Equally important, null findings should be regarded as informative because they help define the limits of neuromodulation rather than simply its successes.

9 Towards personalised neuromodulation

The evidence reviewed here suggests that the field is gradually moving away from the search for universally effective stimulation protocols towards more individualised neuromodulation strategies. Rather than asking which protocol works best on average, an increasingly important question is which intervention is most likely to benefit a given patient, based on their lesion characteristics, residual network architecture, behavioural profile, and treatment-related biomarkers. Recent aphasia-specific evidence supports this direction. Ren et al. (2023), for example, used resting-state functional connectivity to identify an individualized superior frontal gyrus target for theta-burst stimulation and reported greater improvements in WAB-R Aphasia Quotient in both active stimulation groups compared with sham. This study was also highlighted as an exemplar of personalized targeting in the recent meta-analysis by Wei et al. (2026). Similarly, the functionally guided rTMS trial of Li et al. (2026), together with the tract-based findings reported by Ilardi et al. (2026) and Johnson et al. (2026), further suggests that individual differences in structural and functional network organisation may provide clinically relevant information for treatment selection.

However, the identification of an individualized target does not establish that personalization improves clinical outcomes. Clinically responsible personalization requires prospective evidence that biomarker-guided treatment selection provides a meaningful advantage over conventional targeting or other clinically available approaches. This will require adequately powered randomized controlled trials in which the personalization strategy is defined a priori, tested prospectively, and directly compared with a standardized treatment approach. Importantly, such studies should establish not only whether individualized biomarkers predict treatment response, but whether their use provides incremental clinical benefit beyond information that can already be obtained from behavioural assessment, lesion characteristics, or other routinely available clinical measures. The reliability and reproducibility of the proposed biomarkers, their generalizability across patients and centres, and the feasibility and cost of implementing individualized targeting in routine clinical settings will also need to be established.

This transition therefore requires moving from post-hoc identification of predictors of treatment response towards prospectively validated treatment-selection algorithms. Future studies should determine which combination of lesion anatomy, structural and functional connectivity, computational electric-field modelling, and behavioural phenotyping provides sufficient incremental predictive value to guide treatment selection. Evidence from other neurological and neuropsychiatric conditions is beginning to support this framework, including head-to-head comparisons of connectivity-based versus conventional scalp-based targeting (Taylor et al., 2026), network-based approaches to identifying syndrome-specific stimulation targets (Siddiqi et al., 2020), and functional-connectivity-guided targeting within accelerated stimulation protocols (Cole et al., 2020). For post-stroke aphasia, the next step should therefore be to test reproducible and prospectively defined personalization strategies in randomized controlled trials, while determining which patient characteristics and neurobiological markers provide the greatest added value for predicting and optimizing treatment response.

10 Clinical implications

For current practice, four principles are defensible. First, neuromodulation should be considered an adjunct rather than a replacement for speech-language therapy. Second, neither hemisphere nor polarity should be selected from diagnosis alone. Lesion anatomy, recovery phase, residual network function, and the therapy task matter. Third, the expected benefit should be framed modestly: pooled effects are generally small to moderate and may be confined to language domains. Fourth, efficacy should be judged by more than improvement on practised items.

Among available techniques, rTMS currently has the strongest evidence base, whereas conventional tDCS appears easier to implement but yields smaller, more variable effects. Other approaches—including cerebellar stimulation, tsDCS, tACS, and tPBM—remain investigational. Importantly, treatment decisions should not rely solely on aphasia subtype or hemisphere of stimulation. Lesion characteristics, recovery stage, residual network integrity, and the behavioural intervention with which stimulation is paired are likely to be equally important determinants of treatment response. Finally, clinical benefit should be judged not only by improvements on trained language tasks but also by gains in functional communication and everyday participation.

11 Conclusion

Twenty-five years of research have established that neuromodulation can enhance language recovery under specific conditions. The central challenge is no longer to determine whether neuromodulation works, but to understand for whom, when, and through which mechanisms it is most effective. Addressing this challenge will require moving beyond standardised protocols towards biologically informed, network-based, and patient-specific treatment strategies integrated with high-quality speech-language therapy.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. ChatGPT 5.5 (OpenAI, San Francisco, CA, United States) was used for language editing support.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Ryuta Kinno, Showa University Fujigaoka Hospital, Japan

Reviewed by: Thiago Paranhos, University of Texas Southwestern Medical Center, United States

Matteo Varini, University of Turin, Italy

Footnotes

rTMS, repetitive transcranial magnetic stimulation; SLT, speech-language therapy; tACS, transcranial alternating current stimulation; tDCS, transcranial direct current stimulation; TBS, theta-burst stimulation; tPBM, transcranial photobiomodulation; tsDCS, transcutaneous spinal direct current stimulation.