This was a 12-week, prospective, rater-blinded, controlled trial conducted between November 2019 and December 2020 at the Memory Disorder Clinic of Ewha Womans University Mokdong Hospital. Participants were randomly allocated in a 1:1 ratio to two groups: CI using a personal robot (robot) and no CI (control).

This calculation was informed by a prior study involving patients with mild cognitive impairment11 where an electroencephalography (EEG) was used as the primary outcome measure and showed a change in the delta band of μ1-μ0=0.115 over 6 weeks. Given our focus on patients with dementia, we expected similar changes over the 12-week robotic cognitive training period. Using a standard deviation of 0.13 for EEG changes, and aiming for a significance level of 0.05 and a power of 80%, we determined a requirement of 22 participants per group. Adjusting for an expected dropout rate of approximately 15%, we planned to enrol 26 participants in each group, summing up to 52 participants overall.

Fifty-five volunteers aged 60 years or older diagnosed with AD dementia were recruited from the Gwangyang Dementia Center and the Yangcheon Dementia Center. The participants were diagnosed with probable AD dementia based on the National Institute of Neurological and Communicative Disorders and Stroke of the United States and the Alzheimer’s Disease and Related Disorders Association [12].

The inclusion criteria were as follows: 1) age 60-90 years; 2) global clinical dementia rating score of 0.5 or 1; 3) Korean version of the Mini-Mental State Examination-2 (K-MMSE-2) [13] score ≥18; 4) ability to read and write; 5) no visual or hearing impairment severe enough to interfere with cognitive testing/cognitive training; and 6) presence of a reliable informant who live with them or meet them for more than 6 hours a day, at least 3 times a week. The cut-off score of 18 on the K-MMSE-2 was selected because it is widely used in clinical and research settings to define mild AD dementia, reflecting preserved basic functioning with measurable cognitive impairment [14]. Participants were excluded if they had major psychiatric illnesses, a history of cancer in the past 5 years, serious or unstable symptomatic cardiovascular diseases, or other serious medical conditions. In addition, participants who were uncooperative or unable to participate in intervention programs were excluded. Of the 55 participants initially recruited, three were excluded during the screening process because they met one or more of the exclusion criteria.

All patients underwent neuropsychological testing using a standardised neuropsychological battery, the Seoul Neuropsychological Screening Battery 2nd edition, which encompasses five cognitive domains: attention, memory, visuospatial, language, and frontal executive functions [15]. Additionally, laboratory assessments included a complete blood count, blood chemistry, thyroid function tests, syphilis serology, and vitamin B12/folate levels. Brain MRI was performed to confirm the absence of structural lesions including cerebral infarctions, brain tumours, vascular malformations, and hippocampal sclerosis. All participants were being treated with acetylcholinesterase inhibitors and there was no change in the medication dosage during the intervention.

Of the 52 participants, one was excluded because of the withdrawal of consent before randomization. Therefore, the final sample comprised 51 participants (Figure 1), who were randomly allocated into the following two groups: the robot intervention group (robot) (n=27) and the control group (control) (n=24). The research involving human participants was approved by the Institutional Review Board of the Ewha Womans University Mokdong Hospital (EUMC 2019-08-019-001) and conducted in accordance with relevant guidelines and regulations, including the Declaration of Helsinki. Written informed consent was obtained from all participants and their legal guardians.

The study was also registered at the Clinical Research Information Service Site (KCT0004473) on 26/11/2019.

Following the baseline assessments, 51 participants were randomly allocated to one of two groups using simple randomization: the robot intervention group (n=27) and the control group (n=24). Randomization was performed by a single independent researcher.

All participants and their caregivers received 1 hour of education on dementia after undergoing the baseline assessment. This education included the definition, diagnosis, and current treatments of dementia and lifestyle management for dementia, encompassing the importance of cognitive training, healthy foods for the brain, aerobic exercise, and stretching that can be performed at home [7,16].

A robot named Bomy (#Robocare), capable of administering 20 cognitive training programs, was set up at each patient’s home within a week of the baseline evaluation (Figure 2). Participants in the robot group were instructed to engage in cognitive training programs with the robot targeting specific cognitive domains, such as memory, language, visuospatial function, calculation, and frontal executive function (Supplementary Figure 1). They were further required to engage in these programs for a minimum of 60 minutes per day, 5 days a week, over a period of 12 weeks. Each daily training session included at least two different programs: one focused on memory training and the other on various cognitive domains, including language, visuospatial function, calculation, and frontal executive function. The robot was programmed to provide two cognitive training programs at designated times from Monday to Friday. Therefore, over a 12-week period, patients in the robot group underwent a minimum of 120 training sessions, consisting of 60 sessions for memory and 15 sessions each for language, calculation, visuospatial function, and frontal executive function. If the participants did not adhere to the scheduled programs, the robot delivered encouraging messages to promote participation in the cognitive training. If the participant failed to engage in the program after three encouraging messages, the robot would notify the designated caregiver to directly encourage the patient to participate in the cognitive training program. Although the minimum requirement was 5 days per week (Monday to Friday), participants were free to engage in additional training sessions at any time, including weekends, and there was no upper limit on the number of sessions. As a result, some participants performed far more than the minimum requirement.

The primary outcome was change in brain function measured by resting-state EEG, which was measured in eyes-open and eyes-closed conditions for 3 minutes each, with a 19-channel wireless EEG device. EEG parameters were chosen as the primary outcomes because they offer a direct, objective measurement of brain functional changes and can be more sensitive than cognitive test scores in detecting subtle neurophysiological alterations related to dementia progression and intervention effects [17,18]. While cognitive assessments remain essential for clinical evaluation, resting-state EEG provides complementary mechanistic insights and reliable biomarker readouts, thereby strengthening the evaluation of intervention efficacy [19,20].

Resting-state EEGs were recorded over a 3-minute period with eyes open, and another 3-minute period with eyes closed using 19 electrodes based on the international 10-20 system (Fp1, Fp2, F7, F3, Fz, F4, F8, T7, C3, Cz, C4, T8, P7, P3, Pz, P4, P8, O1, and O2) at baseline and at the end of the study. We selected 3 minutes of eyes-closed and artefact-free data based on visual inspection for further analysis. The EEG signals were digitised after filtering with a bandpass of 0.5-70 Hz, and artefacts were removed in two steps. The first step involved the rejection of non-stationary bad epochs, and the second step involved the removal of stationary bad components related to electromyograms, electrooculograms, cardiac signals such as heartbeat, and slow drift waves such as drowsiness to yield cleaned quantitative EEG data using adaptive mixture independent component analysis [21]. At the sensor level, the absolute power of EEG and the square of the amplitudes was calculated using fast Fourier transform spectral analysis in each of the following eight frequency bands: delta (1-4 Hz); theta (4-8 Hz); alpha1 (8-10 Hz); alpha2 (10-12 Hz); beta1 (12-15 Hz); beta2 (15-20 Hz); beta3 (20-30 Hz); and gamma (30-45 Hz). To calculate relative power, the absolute power of each frequency band was divided by the total power. In source-level analysis, standardised low-resolution brain electromagnetic tomography (sLORETA) was used with 68 regions of interest (ROIs) based on the Desikan-Killiany atlas. The imaginary part of the coherence was calculated as the functional connectivity between the 68 ROIs at eight frequencies [22]. All EEG preprocessing, sensor-level data, source-level data calculations, and extractions were performed using a cloud-based AI-driven auto-analysing platform, iSyncBrain® (iMediSync Inc.; http://isyncbrain.com).

The secondary outcomes included the mean scores on three tests based on the Cambridge Neuropsychological Test Automated Battery (CANTAB), which included the spatial working memory (SWM), the paired associates learning (PAL), and the reaction time (RTI) tests (detailed descriptions of the tests are available on the associated website: http://www.cambridgecognition.com/academic/cantabsuite/tests). In brief, the SWM evaluates the visual working memory function to retain spatial information and manipulate remembered items. The error in the SWM task was defined as the time at which the participant revisited a box in which a token had previously been found. The lower the error of the SWM task, the better the performance. The PAL task measured simple visual memory and visuospatial associative learning. The total number of errors from the PAL task was used to measure performance, with lower scores indicating better performance. The RTI task assessed reaction and processing speeds. The outcome measure for the RTI task was the median movement time of the five circles, which referred to the interval between the release of the starting position button and contact with the yellow circle. The raw scores for each of the three tasks were converted to standardized Z-scores using the baseline mean scores and standard deviations of the control group. For better visualisation, the Z-scores were reversed such that positive Z-scores represented better performance.

All primary and secondary outcomes were evaluated by independent researchers blinded to the participants’ intervention status.

Demographic and clinical characteristics were compared using t-tests for continuous variables and chi-square tests for categorical variables. Analysis of covariance was applied to compare baseline neuropsychological tests using age, sex, and years of education as covariates. A generalised linear model was used to estimate group differences in the changes in CANTAB scores before and after the intervention, using age, sex, and years of education as covariates. Significance for all tests was set at α=0.05, two-tailed. All statistical analyses were performed using Stata version 18.0 (Stata Corp).

The robot group showed improved changes in SWM (p=0.039), PAL (p=0.044), and RTI (p=0.033) compared to the control group (Figure 5).

Although a positive correlation was identified between the number of participating sessions and changes in the SWM Zscores, the difference did not reach statistical significance (r=0.414, p=0.068) (Supplementary Figure 2). Further, there were no statistically significant correlations between the number of participating sessions and the changes in PAL (r=-0.073, p=0.770) or changes in RTI (r=0.100, p=0.691).