This was a parallel-group, multicenter, randomized, double-blind, sham-controlled clinical trial of prefrontal tDCS in patients with mild AD, conducted at seven university-affiliated hospitals in Korea. The protocol was approved by local institutional review boards and national regulatory authorities, and written informed consent was obtained from all participants or their legal representatives in accordance with the Declaration of Helsinki. Key inclusion criteria were aligned with contemporary diagnostic standards for AD dementia (1) and included age 55–90 years; a diagnosis of major neurocognitive disorder due to AD according to DSM-5/ICD-10 or equivalent clinical criteria; K-MMSE scores of 18–26; CDR global 0.5–1.0 and CDR-SB 0.5–4.0; stable treatment with acetylcholinesterase inhibitors and/or NMDA receptor antagonists for at least 3 months; and sufficient sensory abilities, with aids as needed, to complete neuropsychological testing. Major exclusion criteria comprised significant neurological or psychiatric comorbidities that could impair cognition, recent alcohol or substance misuse, uncontrolled systemic illness, contraindications to tDCS (e.g. cranial metal implants, skin disease at electrode sites), pregnancy, and participation in other interventional trials. A total of 120 participants were randomized (active tDCS, n = 59; sham, n = 61). The intention-to-treat (ITT) population comprised 106 participants (active n = 53, sham n = 53) with usable baseline and week-26 K-MMSE data for the primary outcome (Figure 1, Table 1).

The target sample size was determined a priori on the basis of the expected between-group difference in change on the K-MMSE, the prespecified primary endpoint at 26 weeks. Previous tDCS trials in patients with mild Alzheimer's disease or related dementias have reported mean between-group differences in MMSE change of approximately 2.5–3.0 points over 6–12 months, with common standard deviations of around four points (9, 10, 15). Assuming a superiority margin of 0.21 points, a common standard deviation of 4.1, a one-sided α of 0.025 (equivalent to a two-sided α of 0.05), and 80% power to detect a between-group difference in 26-week K-MMSE change, we estimated that 88 evaluable patients (44 per group) would be required. Allowing for an anticipated dropout rate of 25%−30% over 6 months, we planned to randomize approximately 118–120 participants (about 59–60 per group). In the present trial, 120 participants were randomized and 106 (53 per group) were included in the ITT population.

Participants were randomly assigned (1:1) to active or sham tDCS using a computer-generated schedule with concealed allocation. Randomization was implemented by an independent statistician or unblinded staff not involved in assessments. The tDCS device was pre-programmed so that the appearance and procedures were identical between active and sham modes, with sham stimulation mimicking the initial ramp-up sensation. Participants, caregivers, clinicians, raters, and primary data analysts remained blind to allocation throughout the trial. A formal assessment of blinding success (e.g., participant/caregiver/rater allocation guess) was not administered.

tDCS was administered with the YMS-201B device (Ybrain Inc., Seongnam, Korea). Following the international 10–20 EEG system, the anode was placed over the left dorsolateral prefrontal cortex (F3) and the cathode over the right dorsolateral prefrontal cortex (F4). Saline-soaked sponge electrodes were used to ensure stable conductivity and participant comfort. Stimulation intensity was set at 2 mA for 30 min per session. Sessions were delivered over the 26-week double-blind period according to a prespecified schedule of five sessions per week, self-administered at home at a convenient (non-fixed) time of day with encouragement to maintain a consistent routine. No standardized concurrent cognitive/language training task was prescribed during stimulation; in-session activities were not systematically recorded, and instructions were identical in the active and sham conditions. Participants were randomly assigned to Active-tDCS or Sham-tDCS. For sham, the device followed the same preparation and placement procedures, but current was discontinued after an initial ~30-s ramp, producing the brief tingling sensation used for blinding while avoiding sustained stimulation (23). To support feasibility and adherence, participants or caregivers received hands-on training in electrode positioning and device operation before treatment initiation. They were instructed to clean the skin and electrodes before/after sessions, record any discomfort or adverse events, and contact the study team with concerns. Regular follow-ups (on-site or remote) were conducted to monitor compliance and reinforce correct placement and procedures, in line with recommendations for remotely supervised/home-based tDCS implementations (21, 22). However, prespecified quantitative protocol-fidelity metrics (e.g., electrode-placement verification counts and remote-supervision frequency) were not systematically recorded for reporting. All device issues and adverse events were documented per protocol.

Neuropsychological and clinical assessments were conducted at baseline and week 26, with several scales additionally administered at week 13; the present report focuses on baseline and week-26 data. The primary outcome was the 26-week change in global cognition measured by K-MMSE. Pre-specified secondary outcomes included clinical severity and function [Clinical Dementia Rating (CDR) and Korean Instrumental Activities of Daily Living (K-IADL)]; memory and visuospatial function [Seoul Verbal Learning Test—Elderly (SVLT-E); Rey Complex Figure Test (RCFT)]; executive function and attention [Korean Stroop Word–Color Test (K-CWST); Digit Span Test (DST); Controlled Oral Word Association Test (COWAT)]; language [Korean Boston Naming Test (K-BNT)]; additional global cognition [Montreal Cognitive Assessment—Korean (MoCA-K)], with totals computed as the sum of item-level scores; and quality of life and behavioral symptoms [Quality of Life in Alzheimer's Disease (QoL-AD), using the dataset's designated total score field; Korean Neuropsychiatric Inventory (K-NPI), total computed as the sum of frequency × severity across domains; and Family Quality of Life–Dementia (FQoL-D), caregiver-reported total score]. For interpretation, lower scores indicated improvement for some measures (e.g., K-NPI, depression scales), whereas higher scores indicated improvement for others (e.g., K-MMSE, K-BNT, QoL-AD, FQoL-D). Safety outcomes included the incidence of any adverse event (AE), serious AEs, treatment-related AEs (defined as events with adverse event relationship (AEREL) codes indicating a definite, probable, or possible relationship to the investigational device), stimulation-site reactions (local symptoms such as pain, discomfort, burning sensation, itching, or skin discoloration at the forehead), and other non–stimulation-site AEs (e.g. headache, diarrhea, anorexia, COVID-19 infection, musculoskeletal pain). Safety events were coded and summarized for the Safety set (n = 118).

The primary estimand was the treatment-policy estimand at week 26, defined as the effect of assignment to active vs. sham tDCS on the primary outcome regardless of adherence, treatment discontinuation, or other intercurrent events. The per-protocol (PP) analysis was prespecified as supportive and targets a while-on-treatment/hypothetical estimand among participants meeting adherence and protocol-compliance criteria. Compliance (adherence) was calculated as the number of completed sessions divided by the prespecified planned sessions, based on device-recorded session completion; the ≥80% threshold was used only to define the PP population. Efficacy analyses were conducted in the ITT and PP populations as defined above; sample size varied slightly by endpoint. K-MMSE was the prespecified primary endpoint and the only outcome for which the trial was powered. All other cognitive domain measures, neuropsychiatric outcomes, and caregiver-reported quality-of-life measures were prespecified as secondary or exploratory outcomes and are interpreted as hypothesis-generating; therefore, p-values for these outcomes are reported nominally without formal multiplicity adjustment. Safety analyses were conducted in the ITT set. For each endpoint and analysis set (ITT, PP), we calculated baseline and week-26 means and standard deviations (SD) by group, the mean change (Δ = week 26—baseline), and SD of change. Between-group differences in change (active – sham) were tested using Welch two-sample t-tests. Two-sided p-values < 0.05 were considered statistically significant at the nominal level. Cohen's d effect sizes were computed as the difference in mean change divided by the pooled SD of change. To support an estimation-focused interpretation for key endpoints, we additionally report covariate-adjusted between-group differences (Active–Sham) with 95% confidence intervals and standardized effects (Hedges' g) for K-MMSE, K-BNT, and FQoL-D in Supplementary Table 4. The primary between-group comparisons at week 26 were based on observed outcomes (complete-case analysis).The primary efficacy analysis was conducted using a complete-case approach, restricted to participants with observed baseline and week-26 outcomes. As sensitivity analyses for missing data, we performed (i) a mixed-effects model for repeated measures (MMRM), which uses all available repeated measures under a missing-at-random assumption, and (ii) multiple imputation for missing week-26 outcomes. In the MMRM, treatment group, visit, and their interaction were included as fixed effects, with study site included as a fixed effect, and an unstructured within-participant covariance was assumed. For safety endpoints and categorical variables (e.g. AEs, sex), between-group comparisons used χ² tests or Fisher's exact tests, as appropriate. In addition, for the primary endpoint we fitted an analysis of covariance (ANCOVA) model with 26-week change in K-MMSE as the dependent variable and treatment group (active vs. sham), baseline K-MMSE, and study site (seven-level factor) as independent variables, and we report the adjusted between-group mean difference with 95% confidence intervals from this model. These covariate-adjusted and longitudinal modeling strategies are widely used in large-sample database studies to improve precision and mitigate bias from baseline imbalance and incomplete follow-up (24, 25).

A total of 120 participants were randomized to active tDCS (n = 59) or sham tDCS (n = 61). According to the disposition dataset, 102 participants (active n = 50, sham n = 52) completed the 26-week double-blind phase, whereas 18 discontinued early. Completion rates were similar between groups (84.7% in the active group and 85.2% in the sham group). The most common reason for dropout was withdrawal of consent by the participant or legal representative [active 7/59 (11.9%), sham 8/61 (13.1%)]; as shown in Figure 1, withdrawals were recorded categorically and more granular reasons for consent withdrawal were not systematically captured for reporting. Additional reasons included loss to follow-up in one sham participant and one death and one investigator-determined discontinuation in the active group. We report (a) trial completion of the double-blind phase (Figure 1) and (b) analysis sets defined by availability of outcome data. The complete-case analysis set was defined by observed baseline and week-26 K-MMSE data, which may include participants who discontinued treatment early but completed outcome assessments. For efficacy analyses, the ITT population included all randomized participants, analyzed according to randomized assignment. A complete-case set was defined as participants with observed baseline and week-26 K-MMSE data (n = 106; active n = 53, sham n = 53) and was used for sensitivity analyses. The per-protocol (PP) population comprised 66 participants (active n = 32, sham n = 34) who completed the 26-week double-blind phase without major protocol violations and with device adherence ≥80% of planned sessions (Figure 1, Tables 1 and 2).

Baseline demographic and clinical characteristics of the ITT population are summarized in Table 1. Mean age was 73.6 ± 7.9 years in the active tDCS group and 73.6 ± 6.3 years in the sham group, and approximately two-thirds of participants were female (69.8% active vs. 67.9% sham). Years of education were comparable between groups (9.7 ± 5.2 vs. 10.2 ± 5.7). Baseline cognitive and clinical severity were typical for mild AD dementia and well-balanced across arms: mean K-MMSE scores were 22.17 ± 2.40 (active) and 22.21 ± 2.56 (sham); mean CDR global scores were 0.53 ± 0.12 and 0.55 ± 0.15, respectively; and mean CDR-SB scores were 2.66 ± 1.22 and 2.75 ± 0.97. Depressive symptoms (SGDS) and neuropsychiatric burden (K-NPI) were modest overall (SGDS 3.81 ± 3.49 vs. 3.68 ± 3.32; K-NPI 5.26 ± 7.96 vs. 6.09 ± 8.86) with no statistically significant between-group differences (all p ≥ 0.47; Table 1). Baseline neuropsychological performance (Table 2) indicated broadly mild impairment across multiple domains (memory, executive/attention, and language), with overall modest neuropsychiatric burden at study entry (K-NPI totals).

Global cognition, as measured by the K-MMSE, declined modestly in both groups over 26 weeks (Table 2). In the ITT population, mean K-MMSE change was −0.53 points in the active tDCS group (22.17 ± 2.40 to 21.64 ± 3.71) and −0.15 points in the sham group (22.21 ± 2.56 to 22.06 ± 4.02); the between-group difference in change was not statistically significant (p = 0.402). MoCA-K scores showed a similar pattern, with small declines in both groups (active Δ −0.68, sham Δ −0.51) and no evidence of a differential treatment effect (p = 0.752). Secondary and exploratory outcomes are presented to characterize domain-specific signals; results are interpreted cautiously in light of multiple comparisons, with emphasis on effect sizes and 95% confidence intervals rather than statistical significance alone.

Domain-specific cognitive outcomes are summarized in Table 2 and Supplementary Table 1. Episodic verbal memory (SVLT-E total recall) declined slightly in both groups (active Δ −0.08, sham Δ −0.38; p = 0.277), and measures of executive function and attention (K-CWST, DST, COWAT) showed only small, non-significant changes with substantial overlap between groups (all p > 0.17). Visuospatial performance on the RCFT showed a numerical divergence between groups. In the ITT set, RCFT scores increased in the active group (mean Δ +1.70) but were essentially stable or slightly lower in the sham group (mean Δ −0.18; p = 0.073). A similar, non-significant trend was observed in the PP population (active Δ +1.77 vs. sham Δ −0.24; p = 0.132; Supplementary Table 2). In contrast, confrontation naming on the K-BNT showed a small but statistically significant between-group difference favoring active tDCS. In the ITT population, the active group was relatively stable or showed a slight improvement (37.45 ± 11.31 to 37.96 ± 10.78; mean Δ +0.51), whereas the sham group exhibited a decline (35.23 ± 13.15 to 32.91 ± 14.00; mean Δ −2.32), yielding a nominally significant between-group difference in change (p = 0.022; Table 2). In the PP set, this pattern was maintained, with active participants again remaining stable or slightly improved (Δ +0.65) and sham participants showing decline (Δ −2.03; p = 0.043; Supplementary Table 2). SGDS and patient-reported quality of life (QoL-AD) demonstrated minimal mean changes in both groups and no meaningful between-group differences (all p ≥ 0.31; Table 2, Supplementary Table 2).

Neuropsychiatric symptoms, indexed by the K-NPI total score, showed numerically smaller worsening in the active tDCS group than in the sham group in the ITT analysis (Table 2). In the ITT population, mean K-NPI change was +0.36 points in the active group and +3.72 points in the sham group (p = 0.065), suggesting a trend toward relatively more stable behavioral symptoms under active tDCS, although the result did not reach conventional significance. In the PP population, however, the direction of effect was reversed: active participants showed a mean increase of +5.03 points, whereas sham participants showed a slight decrease (Δ −0.47; p = 0.028; Supplementary Table 2). Given these discrepant patterns between ITT and PP analyses, the modest sample sizes, and the lack of multiplicity correction, K-NPI findings should be interpreted cautiously as exploratory and analytically unstable rather than as definitive evidence of benefit or harm.

Caregiver-reported family quality of life, assessed by the FQoL-D, showed a consistent pattern favoring the sham group rather than active tDCS (Table 2, Supplementary Tables 1, 2). In the ITT population, mean FQoL-D scores declined in the active group (91.83 ± 7.93 to 89.64 ± 13.74; mean Δ −2.19) but improved in the sham group (94.94 ± 9.54 to 96.89 ± 9.82; mean Δ +1.94), resulting in a nominally significant between-group difference in change (p = 0.043; Table 2). In the PP set, this pattern was even more pronounced: active participants showed a mean decline of −2.97 points, whereas sham participants improved by +3.18 points (p = 0.042; Supplementary Table 2). Item-level analyses (Supplementary Table 3) indicated that this divergence was not driven by a single question but by small, consistent differences across multiple items related to emotional support, access to medical care, time for caregivers' own activities, and availability of help in special situations. For example, sham caregivers reported improvements in “having time for their own hobbies” (item 8; Δ +0.09 vs. −0.28, p = 0.016) and “receiving satisfactory emotional support” (item 10; Δ +0.15 vs. −0.23, p = 0.013), whereas active caregivers tended to report stable or slightly worse scores on these items.

Sensitivity analyses using an MMRM and multiple imputation for missing week-26 K-MMSE yielded results consistent with the primary complete-case analysis; no conclusions regarding statistical significance were changed (Supplementary Table 1).

Adverse events during the 26-week double-blind period are summarized in Table 3. Any AE occurred in 33/53 participants (62.3%) in both sham and active groups. Treatment-related AEs, all of which were stimulation-site reactions at the forehead, were reported in 24/53 (45.3%) sham participants and 31/53 (58.5%) active participants (p = 0.243). Stimulation-site reactions included pain, discomfort, burning sensation, itching, and skin discoloration; these events were generally mild and transient. Other (non–stimulation-site) AEs such as headache, diarrhea, anorexia, COVID-19 infection, and musculoskeletal pain—occurred in ≤ 2 participants ( ≤ 3.8%) per event type and were numerically more frequent in the sham group [13/53 (24.5%)] than in the active group [6/53 (11.3%), p = 0.127]. Serious AEs were infrequent: they occurred in 3/53 (5.7%) sham participants and in none of the active participants (p = 0.243). No serious AE was judged to be related to the investigational device. Detailed reasons for dropout are summarized in Table 4.