This was a prospective, single-center, rater-blind, randomized controlled trial, which was conducted from November 21, 2019, to July 13, 2020, at the Memory Disorder Clinic of Ewha Womans University Mokdong Hospital in Seoul, Republic of Korea. This study was reported in accordance with the CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile Health Applications and Online Telehealth) checklist (Multimedia Appendix 1). Patients and the public were not involved in the design, conduct, reporting, or dissemination plans of this trial.

A total of 81 CU community-dwelling individuals were recruited from the Ewha Womans University Mokdong Hospital in Seoul. All participants underwent a comprehensive neuropsychological assessment using the Seoul Neuropsychological Screening Battery, second edition (SNSB-II) [28] by trained clinical neuropsychologists. CU was defined as those aged 60 or above who scored more than –1 SD below the mean of age- and education-matched norms in tests evaluating memory, attention, language, visuospatial, and frontal executive functions in the SNSB-II. We excluded individuals with any of the following characteristics: (1) suspected or confirmed mild cognitive impairment or dementia; (2) suspected or confirmed major neurological or psychiatric illnesses, including major depressive disorders; (3) contraindications to MRI; (4) visual or hearing impairments severe enough to interfere with questionnaire response; (5) history of medications that could affect cognitive and emotional functions in the past 3 months; and (6) other major medical problems. Of the 81 participants, nine were excluded because of the withdrawal of consent before the baseline assessment. Therefore, a total of 72 participants were included in this study.

For the formal sample size calculation, we used the cortical thickness effect-size estimate specified in the study protocol: an expected longitudinal cortical thickness change of 0.046 with an SD of 0.053, corresponding to a standardized effect size of Cohen d=0.87 [29]. The sample size calculation was based on a 2-tailed independent-samples t test comparing mean longitudinal change between the 2 groups. Assuming 2 parallel groups, a 2-tailed α of .05, and 90% statistical power to detect group differences, the required sample size was estimated to be 30 participants per group. To accommodate an anticipated dropout rate of approximately 20%—commonly observed in MRI-based intervention studies—we planned to enroll 36 participants per group, yielding a total target sample of 72 participants.

Written informed consent was obtained from all participants. The study protocol was approved by the Institutional Review Board (IRB) of Ewha Womans University Mokdong Hospital (IRB number 2019-06-028). The study was conducted in accordance with the Declaration of Helsinki and relevant guidelines and regulations. Participant data were handled confidentially and deidentified for analysis in accordance with applicable institutional policies. Participants did not receive financial compensation for participation. This trial was registered with the Clinical Research Information Service (CRIS), Republic of Korea (KCT0004630).

After the baseline assessment, 72 participants were randomly allocated in a 1:1 ratio to 2 groups: the intervention group (n=37) and the control group (n=35). A simple, computer-generated randomization sequence was created by an independent statistician who had no contact with study participants. To ensure allocation concealment, the sequence was stored in a password-protected file and revealed only after the baseline assessment was finalized. Trained research staff enrolled participants and had no access to the allocation sequence. Outcome assessors were blinded to group assignments throughout the study.

All outcome evaluators were blinded to the group assignment of participants and were not involved in the intervention.

All participants from both groups received 30 minutes of education about dementia prevention after undergoing the baseline assessment. Participants in the control group received education only and were instructed to continue their usual activities for 12 weeks without any additional cognitive training. This education included the definition, diagnosis, and current treatments of dementia and lifestyle modifications for dementia prevention, encompassing the importance of cognitive training, healthy foods for the brain, aerobic exercise, and stretching that could be done at home [30,31]. Beyond the study intervention, no other cointerventions were provided. To avoid confounding effects, participants continued their routine health care and agreed not to participate in any other cognitive enhancement programs throughout the trial.

All participants in the intervention group used Saemi (HAII Corporation), a chatbot-based cognitive training program delivered via a mobile messaging platform. The program provides representative brain-training tasks grouped by cognitive domains—language, memory, attention, calculation, and executive function—as illustrated in Figure 1. Participants were instructed to complete training for at least 30 minutes per day, 5 days per week, for 12 weeks. Each daily session comprised at least 2 tasks, including one consistently focused on memory and another targeting a different domain (language, visuospatial function, calculation, or executive function). Actual intervention exposure was quantified using app usage logs. Usage metrics included total training time, completed session counts, active days, and mean use time per active day. Time-based adherence rate was calculated as the percentage of the recommended minimum total training time completed according to app usage logs. The prescribed minimum dose was 30 minutes per day on 5 days per week for 12 weeks, corresponding to 30 total hours across the intervention period. Adverse events were monitored throughout the trial via participant self-reports and spontaneous reporting.

The figure shows example tasks targeting language, memory, attention, calculation, and executive function in a chat-based mobile format.

Statistical analyses were performed using STATA software (version 18; StataCorp). Baseline demographic and clinical characteristics were compared between groups using the chi-square test for categorical variables, independent-samples t tests for continuous variables, and the Mann-Whitney U test for nonnormally distributed variables. Baseline neuropsychological test performances were compared using analysis of covariance adjusted for age, sex, and years of education.

For cognitive outcomes, raw scores from each task were transformed into standardized z scores using the mean and SD of the control group at baseline; where necessary, scores were oriented such that higher z scores indicated better performance [47].

The main analytical approach for longitudinal changes in DTI metrics and CANTAB outcomes was a linear mixed-effects model, including fixed effects for group, visit (baseline vs follow-up), and the group-by-visit interaction, with a participant-specific random intercept to account for within-subject correlations. Models were adjusted for age, sex, years of education, and the baseline value of each outcome. Missing outcome data were handled implicitly using maximum likelihood estimation in the mixed-effects framework; inference relies on the standard missing-at-random assumption [48].

The main analysis followed the intention-to-treat (ITT) principle, including all randomized participants in the groups to which they were assigned, regardless of adherence. In addition, a supplementary per-protocol (PP) analysis included only those participants who completed the intervention and the follow-up assessment [49].

ROI diffusion metrics were analyzed using the same mixed-effects framework. To account for multiple comparisons across ROIs, FDR correction was applied, with Bonferroni-adjusted and uncorrected P values reported for descriptive purposes.

Associations between changes in WM microstructure (DTI metrics) and cognitive performance (CANTAB z scores) were examined using partial correlation analyses, adjusting for age, sex, years of education, and baseline imaging and cognitive scores. For these exploratory structure-function analyses, missing values were handled using multiple imputation by chained equations with predictive mean matching (m=20 imputations; 5 iterations) under a missing-at-random assumption, implemented in R (version 4.4.1; mice package) [50]. Imputed datasets were analyzed separately and combined using standard multiple-imputation pooling rules [51]. Imputation diagnostics were evaluated using density plots. P values were adjusted for multiple comparisons using the Benjamini-Hochberg FDR procedure.

This trial was prospectively registered with the CRIS of the Republic of Korea (KCT0004630), with cortical thickness and resting-state functional magnetic resonance imaging (fMRI) specified as the registered primary imaging outcomes. These modalities were acquired as planned; however, modality-specific longitudinal analyses of cortical thickness and resting-state fMRI have not yet been completed and are not included in the present report. DTI was included from the outset in the IRB-approved neuroimaging protocol but was not separately listed as an outcome in the CRIS registration. Because DTI processing and longitudinal analyses had been completed for the cohort, this study focuses on longitudinal changes in DTI-derived WM microstructure together with cognitive outcomes and should be interpreted as a focused analysis rather than as a report of the registered primary imaging outcomes. Separately, the registered protocol described a 60-minute daily cognitive training schedule, whereas the implemented intervention instructed participants to complete at least 30 minutes per day, 5 days per week for 12 weeks. This change was made before participant enrollment to reduce participant burden and improve the feasibility of sustained home-based training in older adults.

Of the 72 participants who underwent baseline assessment, 9 (12.5%) participants withdrew from the study: 5 (13.5%) from the intervention group and 4 (11.4%) from the control group. Consequently, 63 (87.5%) participants completed the final assessment (Figure 2). No intervention-related adverse events were reported in either group. Baseline demographic characteristics, clinical variables, neuroimaging characteristics, and cognitive performance across attention, visuospatial function, memory, language, and frontal executive domains did not differ significantly between the intervention and control groups (Table 1).

At week 12, the intervention group showed significantly less reduction in FA than the control group in the right tapetum, left tapetum, and right PTR after FDR correction (P=.03; P=.03; and P=.04, respectively; Table 2; Figure 3). Detailed longitudinal values for each group are provided in Table S1, and within-group pre-post comparisons based on paired t tests are provided in Table S4 in Multimedia Appendix 2.

Regarding MD, a nominal group difference was observed in the left cingulum (cingulate gyrus; P=.04), but this finding did not remain significant following correction for multiple comparisons (FDR-adjusted P=.37; Table S5 in Multimedia Appendix 2).

In the PP analyses (participants who completed the follow-up assessment), the results were consistent with the ITT findings. Participants who adhered to the intervention protocol demonstrated less reduction in FA compared with controls at week 12 in the right tapetum (–0.03 for intervention vs –0.06 for control group; FDR-adjusted P=.04) and right PTR (–0.02 vs –0.03; FDR-adjusted P=.04; Table S6 in Multimedia Appendix 2).

Participants used the program for a mean of 30.73 (SD 12.84) minutes per active day and were active on a mean of 64.24 (SD 16.65) days over the 12-week intervention period, corresponding to approximately 5.35 (SD 1.39) active days per week. Participants completed a mean of 372.52 (SD 200.44) in-app training sessions. The mean total training time was 33.62 (SD 18.53) hours, and the mean time-based adherence rate was 112.1% (SD 61.8%), indicating that participants, on average, exceeded the prescribed minimum training dose of 30 total hours across 12 weeks. In unadjusted exploratory association analyses within the intervention group, the number of active days was positively correlated with PAL z score improvement (r=.374; P=.04), suggesting that more frequent engagement with the intervention was associated with greater memory improvement. No significant associations were observed between app-use metrics and FA changes.

This study demonstrated that a 12-week chatbot-based cognitive training intervention was associated with attenuated decline in FA within selected WM tracts and improved episodic memory in CU older adults, a pattern broadly consistent with a brain maintenance framework.

One of the major findings of this study was that DTI analysis revealed attenuated decline in FA in the intervention group compared with controls, most notably in the right PTR and the bilateral tapetum. Notably, significant group-by-time effects were observed primarily in FA, whereas complementary diffusivity metrics (MD, AD, and RD) showed largely null findings after multiple-comparison correction. FA is a composite measure influenced by fiber coherence, myelination, and axonal properties; thus, the present results are best interpreted as relative maintenance of microstructural organization rather than definitive structural restoration. Given the 12-week intervention duration and the modest effect sizes typically observed in diffusion metrics, detectable changes in MD/RD/AD may require longer follow-up and/or larger samples. These tracts are major long-range projection and commissural pathways that support interregional communication and have been implicated in cognitive aging [22,52]. While WM changes in older adults are often widespread, selective differences in projection and commissural fibers observed here are consistent with previous work suggesting activity-related modulation of myelinated pathways [53].

Projection fibers such as the PTR support posterior thalamo-cortical integration, which is relevant to visuospatial information processing and broader cognitive functions [52]. Given that PAL involves visuospatial associative binding [32,34], the observed attenuated FA decline in the right PTR may reflect relative maintenance of long-range pathways supporting visuospatial associative memory. Similarly, commissural fibers such as the tapetum showed attenuated FA decline, reinforcing their role in interhemispheric communication and integration. These commissural fibers support efficient interhemispheric information transfer, which is relevant to memory and higher-order cognition—functions that are particularly vulnerable to age-related decline [54-56].

Furthermore, the intervention group showed significant improvement in PAL performance, a task particularly sensitive to episodic memory. This domain‑specific response aligns with the design of the Saemi program, in which memory training was a core, recurring element. The selective improvement in memory is consistent with previous studies reporting task‑specific benefits of targeted cognitive interventions [8]. Because PAL is a visuospatial associative memory task, the observed improvement may reflect both a memory-related benefit and alignment between the intervention content and the specific cognitive demands of the task. Accordingly, these findings should not be interpreted as evidence of broad cognitive transfer across multiple domains. In contrast, no significant intervention-related changes were observed in the other cognitive outcomes, including attention, processing speed, executive function, and visuospatial recognition memory. This pattern suggests that the intervention effect on cognition was selective to episodic memory rather than generalized across all measured cognitive domains. The nonsignificant findings in the other cognitive outcomes may reflect the memory-focused structure of the intervention, limited transfer to less directly trained domains, measurement sensitivity, the relatively short intervention duration, or limited statistical power to detect smaller effects.

These findings are more consistent with brain maintenance reflected by attenuated microstructural decline than with newly induced structural restoration. Because diffusion metrics provide an indirect index of WM microstructure and do not directly quantify myelin or synapse formation, these findings should not be interpreted as evidence of structural “restoration.” Nevertheless, repeated engagement in cognitively demanding tasks has been linked to experience-dependent neuroplastic processes—including myelin-related changes and synaptic remodeling—in prior work [27,57,58], which posits that relative maintenance of brain integrity supports sustained cognitive function in older adults [21]. Although these findings are conceptually aligned with a brain-maintenance framework, the absence of longer-term follow-up assessments precludes definitive conclusions regarding the durability of this mechanism. In summary, the observed pattern is more consistent with relative maintenance of WM microstructural organization in pathways relevant to cognitive integration and memory processing, which may be related to intervention-associated improvements in episodic memory [46,53] and is conceptually aligned with a brain maintenance framework [21].

Building on the observed intervention effects on both cognitive performance and WM microstructure, we explored whether changes in these domains were interrelated. Using the imputed ITT dataset, exploratory partial correlation analyses were performed across all participants to examine whether changes in PAL performance were associated with changes in WM microstructure. Change scores were calculated as follow-up minus baseline values. Positive PAL z score changes indicate improvement in performance, whereas positive FA changes indicate relative preservation or increase in FA. Partial correlation analyses suggested an association between changes in episodic memory performance and WM microstructural integrity in task-relevant pathways. In addition, PAL improvement was also associated with FA change in the right SS, an associative pathway that includes major posterior cortico-cortical fibers (eg, inferior fronto-occipital and inferior longitudinal components) supporting visual information transfer and integration. Given that PAL relies on visuospatial associative binding, the concurrent PTR-SS associations may suggest that relative maintenance of posterior projection and associative pathways is related to memory-related performance during aging. However, these exploratory structure-function correlations cannot establish causal or mechanistic relationships and should be interpreted as hypothesis-generating findings that require confirmation in larger studies with longer-term follow-up. Nevertheless, this pattern is broadly consistent with the brain maintenance framework, which emphasizes relative preservation of neural structure and function in successful cognitive aging [21,22,52].

Taken together, these findings suggest a possible structure-function relationship in which improvements in episodic memory were associated with relative maintenance of white matter microstructural organization, consistent with a brain maintenance interpretation. Rather than demonstrating a definitive mechanism, our findings indicate that better episodic memory performance was associated with relatively preserved WM pathways implicated in memory-related processing. This interpretation is consistent with the integrated maintenance model, which emphasizes the contribution of both cognitive and structural resilience to successful cognitive aging [21,53,59,60].

Several limitations should be acknowledged. First, the relatively small sample size may have limited the statistical power to detect smaller effect sizes and reduced the generalizability of the findings. Future studies with larger, more diverse samples are needed to confirm these results. Second, DTI-derived outcomes were not listed as outcomes in the trial registration, although DTI was included in the IRB-approved neuroimaging protocol. Therefore, these findings should be interpreted as a focused analysis rather than as the primary confirmatory findings of the registered trial. Third, although previous studies suggest that cognitive training effects can persist long-term [1,61], this study assessed only the immediate postintervention effects of a 12-week program. Therefore, the durability and longitudinal trajectory of the observed cognitive and WM changes could not be determined. In addition, without follow-up assessments beyond the intervention period, it remains unclear whether the observed WM differences reflect sustained neuroplastic change or more transient fluctuations. Longitudinal follow-up studies with repeated assessments over time are necessary to clarify the durability of these effects and the potential need for booster sessions. Fourth, potentially influential lifestyle factors such as sleep, physical activity, and social engagement were not assessed. These factors are known to affect WM integrity and may interact with cognitive training effects. Future research could consider multimodal interventions that integrate cognitive, behavioral, and lifestyle components [62]. Fifth, although the intervention emphasized memory training, likely contributing to the significant PAL improvements, no effects were observed in other cognitive domains. This domain-specific pattern suggests that the cognitive benefit was not broadly generalized across domains. Because PAL is a visuospatial associative memory task, the observed improvement may reflect both a memory-related benefit and alignment between the training content and the specific task demands. Future iterations of the program could adjust the training balance or incorporate more targeted modules to evaluate broader transfer effects. Another limitation is that the study did not include an active control condition. Accordingly, nonspecific effects such as expectancy, engagement, general mental stimulation, or digital interaction cannot be fully excluded. Therefore, the observed effects should be interpreted cautiously and should not be attributed solely to the specific cognitive-training content of the intervention. Future studies should include an active control condition matched for contact time, digital interaction, and engagement to better distinguish the specific effects of chatbot-based cognitive training from nonspecific intervention-related effects.

Nonetheless, this study demonstrated that a 12-week chatbot-based cognitive intervention was associated with improved episodic memory and attenuated decline in WM microstructural integrity within key neural pathways. These findings provide early evidence that chatbot-based cognitive training was associated with improvements in episodic memory and attenuated decline in WM microstructural integrity, suggesting potential cognitive and neurostructural effects that warrant confirmation in larger trials.