Authors: Suebeen Kim, Doo Young Kim, Si-Woon Park, Namo Jeon, Taeksoo Jeong, Min-Soo Kang, Sangwook Park
Categories: Original Article, Artificial intelligence, Cognitive dysfunction, Mobile applications, Stroke, Telerehabilitation
Source: Annals of Rehabilitation Medicine
Doi: 10.5535/arm.250060
Authors: Suebeen Kim, Doo Young Kim, Si-Woon Park, Namo Jeon, Taeksoo Jeong, Min-Soo Kang, Sangwook Park
To test the feasibility and usability of an artificial intelligence (AI)-guided mobile cognitive telerehabilitation program for patients with stroke or older adults with mild cognitive impairment (MCI).
Thirteen participants with cognitive impairment (Mini-Mental State Examination [MMSE] score≤26; nine with stroke and four with MCI) were enrolled in the study. Each participant was provided with an AI-guided mobile cognitive rehabilitation program (Zenicog^®^). Participants were instructed to complete 24 sessions within 6 weeks, and those with sufficient adherence (≥70%, 17 sessions) were included in the analysis. Cognitive assessments included the MMSE, digit span, and Trail Making Tests A & B. The usability questionnaire investigated equitable use and flexibility in use, simple and intuitive use, perceptible information, tolerance for error, low physical effort, size and space for use, overall product quality, overall satisfaction.
Eleven participants completed the study, and 10 participants met adherence criteria. The MMSE score increased significantly from 24.00 [21.00, 25.75] at baseline to 27.50 [26.00, 28.75] after intervention. The overall product quality (Likert 1–5) score was 4.00±0.87. The lowest score in the usability questionnaire was for tolerance for error. Female participants and participants with <12 years’ education gave lower scores for tolerance for error and equitable/flexibility in use, respectively.
The AI-guided mobile cognitive telerehabilitation program is feasible and potentially beneficial for improving cognitive function in patients with stroke or older adults with MCI. Individuals who are less familiar with electronic devices require special consideration to improve their usability.
Cognitive impairment, frequently emerging in older adults post-stroke or as part of conditions such as mild cognitive impairment (MCI), substantially hampers daily functioning and reduces quality of life [1,2]. In such cases, traditional rehabilitation methods, requiring frequent visits to healthcare facilities, pose challenges for those with physical or logistical constraints. Over time, telerehabilitation, which leverages remote therapy, has been developed as a means of providing traditional rehabilitation, offering several advantages over earlier approaches. The evolution of telerehabilitation offers an adaptable solution to overcome the abovementioned barriers, providing the necessary therapeutic interventions within the comfort of an individual’s home [3]. However, telerehabilitation still has several limitations. For instance, the restricted human interaction requires more time to establish a therapeutic relationship between therapists and patients. This makes it challenging for therapists to assess patients and to adjust exercises as needed.
A further advancement for addressing these limitations has been the development of computerized cognitive rehabilitation (CCR). CCR has become a pivotal method for enhancing cognitive functions, such as visual response, attention, and problem-solving, through interactive software [4]. This approach not only facilitates independent patient learning but also allows for precise monitoring of therapeutic outcomes. Despite its benefits, this approach still relies on therapists to manage patient engagement and adjust difficulty levels, highlighting the need for more autonomous systems [5], such as integration of artificial intelligence (AI).
The incorporation of AI into mobile health applications presents a transformative approach to cognitive rehabilitation [6]. The ability of AI to customize treatment plans based on real-time patient data promises to increase intervention efficacy [7]. By deploying these AI-enhanced programs on devices, the technology can meet specific patient needs, promoting greater adherence and engagement [8].
Zenicog^®^ (Mindhub Inc.), a state-of-the-art CCR program tailored to stroke survivors and older adults with MCI, utilizes advanced AI and cloud technologies to address the limitations of current therapies. Its design allows for remote monitoring of patient engagement and adaptive difficulty adjustments based on ongoing performance assessments. This minimizes the need for direct therapist intervention and enables effective, standardized therapy in a home setting. Even without real-time intervention by healthcare providers, activity logs allow for periodic monitoring of patient compliance during outpatient rehabilitation sessions, whether conducted weekly or at longer intervals. AI-driven difficulty adjustments overcome the challenges posed by limited therapist interaction in remote rehabilitation, reducing errors and improving patient adherence. Remote access further enables therapists to review and adjust treatment as needed.
The present study investigated the feasibility and usability of Zenicog^®^ by evaluating the participants’ cognitive function to test the therapeutic potential of this innovative AI-based approach. We further evaluated the program’s user-friendliness (user satisfaction and accessibility) to ensure widespread acceptance and efficacy in the target population.
This was a multi-center single-group pre–post study without a control group, intended as a preliminary study to assess feasibility. It was conducted using a case series design with a pre–post comparison to evaluate the effectiveness and usability of this AI-guided mobile cognitive telerehabilitation program.
The sample size of this study was determined based on methodological recommendations for pilot and feasibility trials. Assuming the medium to large effect size inferred from the previous study [9], a sample size of 10 participants is generally sufficient to achieve 80 percent power [10], and this criterion was used to determine the sample size in the present study.
Participants were enrolled if they exhibited cognitive impairment, as indicated by a Mini-Mental State Examination (MMSE) score of 26 or less, or if they were stroke patients in the subacute phase. This phase encompassed those with less than 6 months post-stroke onset and included individuals with both hemorrhagic and ischemic strokes.
Exclusion criteria included visual or hearing impairments limiting the use of computerized devices and inadequate digital literacy to operate smart devices smoothly. Individuals with severe mental illnesses, such as schizophrenia, dissociative disorders, or major depressive disorders (primary or secondary), were excluded. Additionally, participants taking certain medications, including antipsychotics, antidepressants, drugs used for treating alcohol addiction, sedatives (e.g. tranquilizers, calmatives, benzodiazepines), or opioids (e.g., oxycodone), were not eligible for participation.
A novel AI-guided mobile cognitive rehabilitation program (Zenicog^®^), which consists of attention, memory, and executive training contents, was utilized by participants to improve cognitive function. The specific functions of the Zenicog^®^ system are as follows. It provides a total of 61 training contents classified into seven domains, and offers personalized training contents with optimized sessions, sets, and repetitions for efficient training by AI-based cloud system. It can provide data on usage period, number of training sessions, training outcomes, and time of use, thereby supporting prescription and patient management.
Participants were provided with a tablet PC on which the Zenicog^®^ mobile cognitive rehabilitation software had been installed (Fig. 1). Participants were instructed to complete at least 24 sessions (30 minutes per session) within 6 weeks. Adherence criterion was applied to participants who completed the study and was defined as attending more than 70% (≥17 of the 24 scheduled sessions), which was considered sufficient participation for inclusion in the analysis. The sessions were performed independently by the participants without the supervision of a therapist and were aided by AI. Patients with stroke continued with their conventional rehabilitation treatments alongside this intervention.
The specific software and hardware requirements for the implementation of Zenicog^®^ are presented in Supplementary Material S1. The Zenicog^®^ system applied standardized security measures, including OAuth 2.0–based token authentication for user login, SSL (https) encryption for data transmission, and secure RESTful API and socket communication.
An AI-based recommendation system was integrated into the program to dynamically adapt training plans in response to each participant’s ongoing performance. The AI-based personalized rehabilitation training recommendation system consisted of four key modules.
(1) Training record collection stored data such as accuracy and response time to monitor patient progress.
(2) Training task analysis drew on a categorized problem set defined by difficulty and time constraints to match tasks with patient performance.
(3) Training performance prediction estimated future outcomes based on prior results.
(4) Training recommendation used optimization algorithms to suggest suitable tasks and adapt session content in real time.
The recommendation algorithm applied in this study did not rely on big data or deep learning approaches. Instead, it selected tasks from a structured item bank, tailoring recommendations to real-time performance indicators such as difficulty level and time restrictions. This cloud-enabled AI rehabilitation system also allowed researchers to oversee participation and training trends remotely, generating useful data to guide future evidence-based decisions in cognitive rehabilitation. (Fig. 2)
The Zenicog^®^ software provides detailed training logs, such as the number of training sessions completed, mean training time (minutes), difficulty level, number of problems attempted and correct/incorrect responses, and mean/domain-specific accuracy rate for attention, memory, and executive function.
Primary outcomes included changes in cognitive function, measured by MMSE 2, Digit Span Forward (DSF), Digit Span Backward (DSB), and Trail Making Tests A & B (TMT-A & -B). Secondary outcomes included quality of life indicators (5-level EuroQol Group-5D [EQ-5D-5L] and EuroQol Group visual analog scale), self-esteem (Self-Efficacy Scale, SES), depression (Korean version of the Center for Epidemiological Studies–Depression Scale [K-CES-D]), activities of daily living (Seoul–Instrumental Activities of Daily Living and modified Barthel Index), and usability assessed through a structured questionnaire.
The MMSE was administered using the Korean version of the MMSE 2, a screening test designed to assess cognitive function through a series of questions. The maximum score is 30, with higher scores indicating better cognitive function [11,12]. The Digit Span Test is a neuropsychological assessment where participants repeat a sequence of numbers provided by the examiner. It is used to evaluate cognitive functions, including attention and working memory. Lower scores indicate a decline in cognitive ability, as individuals recall fewer numbers [13]. TMT-A & -B are widely used neuropsychological assessments for evaluating various cognitive functions. Performance is measured by the time (in seconds) taken to complete the task, with longer completion times indicating impaired cognitive function [14].
Participants were instructed to self-report any adverse events during the intervention period. Potential events were also checked at the end of the study through a usability questionnaire.
The user survey (Supplementary Material S2) was developed with reference to the principles of Universal Design (UD) as outlined by Pramuka and van Roosmalen [15]. The seven UD principles—equitability of use, flexibility of use, simplicity and intuitiveness of use, perceptibility of information, tolerance of error, minimization of physical effort, and suitability of size and space for approach and use—developed by researchers at North Carolina State University, were utilized to address various issues that may hinder the effectiveness of rehabilitation provided by this approach. The UD principles served as a framework for selecting questions to evaluate user experience and accessibility and guided the design of the survey to evaluate potential challenges that stroke patients might encounter when using the remote rehabilitation program.
For example, UD Principle 3 (simplicity and intuitiveness of use) was applied to formulate questions such
Are the instructions for operation clearly explained?
Is the layout of elements on the screen easy to understand?
Are important elements arranged so as to be clearly visible?
Is it more convenient than previously used devices?
The survey items were reviewed by the authors to evaluate their relevance and appropriateness. A consensus process was conducted using the modified Delphi method, which facilitated the prioritization and selection of suitable items for inclusion.
Participants responded to the items in the survey by using a 5-point Likert scale, ranging from 1 (poor) to 5 (excellent). System usability was measured using a structured 25-item questionnaire covering eight key domains, including equitable use and flexibility in use, simple and intuitive use, perceptible information, tolerance for error, low physical effort, size and space for use, overall product quality, and overall satisfaction.
Statistical analysis was performed using SPSS software (version 22.0; IBM Corp.). Paired t-tests were used to compare pre- and post-intervention MMSE scores and usability ratings. A p-value<0.05 was considered statistically significant.
This study was conducted in accordance with the Declaration of Helsinki. Ethics approval for this study was granted by the Institutional Review Board of the International St. Mary's Hospital (IS23OISE0062, IS23OISE0063) and Public Institutional Review Board of the Korea National Institute for Bioethics Policy (P01-202312-01-025). All participants provided informed consent prior to their inclusion in the study. Consent for publication is not applicable. Additionally, this trial was registered at the Clinical Research Information Service (CRIS) under the identifiers KCT0008968 and KCT0008969.
Thirteen participants with cognitive impairment were enrolled, including nine stroke patients in the subacute phase and four participants with MCI. The study completion rate was 84.6%, with 11 of the 13 enrolled participants completing the study.
Two participants withdrew after enrollment owing to lack of motivation; one from the MCI group, before the initial assessment, and the other from the stroke group, after six sessions.
The overall session adherence rate was 90.9%, with 10 out of 11 participants who completed the study meeting the prespecified adherence criterion (≥17 of 24 sessions). Therefore, outcome analyses were conducted in 10 participants who achieved session adherence, while usability analyses included all 11 participants who completed the study.
Among the 10 participants who achieved the adherence criterion, the total training time was 1,096±627.14 minutes, and the mean training time per session was 43.59±20.98 minutes, exceeding the planned 30-minute target. On average, participants completed 24.7±2.63 total sessions.
The participants had an average age of 67.7±8.7 years, and seven were females (70.0%). The average duration of education was 11.2±3.0 years. The stroke participants were on average 67.7±43.2 days post-onset. Among the seven participants with stroke, three had infarction and four had intracerebral hemorrhage. Lesion sites were located in the supratentorial region in five participants and in the infratentorial region in two. Lesion laterality was on the right side in three participants and on the left side in four. Cardiovascular comorbidities were present in eight participants (one in MCI, seven in stroke), and one participant had quadriplegia. General characteristics and a comparison between the MCI and stroke groups are presented in Table 1.
The MMSE score significantly increased from 24.00 [21.00, 25.75] at baseline to 27.50 [26.00, 28.75] after the intervention (p=0.007). Other cognitive measures, such as DSF, DSB, and TMT-A & -B scores, showed improvement after intervention but without statistical significance. Health-related quality of life, self-efficacy, and depression did not show significant changes after the intervention (Table 2). The mean accuracy rates and mean reaction times for the memory, attention, and executive function domains are summarized in Table 3. No device-related adverse events or discontinuations due to device-related side effects were reported.
The overall product quality and overall participant satisfaction in the usability questionnaire were 4.00±0.87 and 3.94±1.18, respectively (Table 4). The domain that showed the lowest score was the tolerance of error (3.14±1.1). Female participants gave lower scores for the tolerance of error than did the male participants (Table 5). Moreover, participants with fewer than 12 years of education gave lower scores for equitable and flexibility of use than did those with 12 or more years of education (Table 5). Participant feedback highlighted areas for improvement in usability, including difficulties with error correction, small font sizes, unclear pronunciations, and physical discomfort; this was particularly true for users with impairments such as hemiparesis. Structured questionnaire results showed that while the program was generally well-received, concerns remained about error tolerance and navigation complexity.
This feasibility study explored the use of an AI-guided mobile cognitive telerehabilitation program, Zenicog^®^, tailored to individuals with stroke and MCI. We conducted an exploratory analysis to provide a more comprehensive understanding of participants’ engagement in the intervention. The completion rate, referring to the proportion of enrolled participants who completed the final assessment (≥75%), was 84.6%, as 11 of the 13 initially enrolled participants completed the study. Two participants dropped out during the study. In-depth interviews revealed that their withdrawal was not due to dissatisfaction with the study protocol or adverse effects of the software, but rather due to a lack of motivation, as they preferred to focus on rehabilitation targeting physical disabilities. The remaining 11 participants completed the protocol and reported no significant adverse effects from participation. The adherence criterion, which required participants to complete at least 70% of the planned sessions (≥17 of 24), was met by 10 of the 11 participants who did not discontinue the study prematurely. One participant did not meet the adherence criteria due to a mild recurrence of stroke, which limited the ability to complete the planned number of sessions. This issue was unrelated to the device or study procedures. Accordingly, this participant was excluded from outcome analyses, although their usability responses did not differ substantially from those of other participants. Throughout the study, no device-related serious adverse events were reported. The mean usability score was ≥3.5 on a 5-point scale, meeting the predefined threshold.
The results indicated a statistically significant improvement in MMSE scores post-intervention, underscoring the potential of AI-integrated approaches in enhancing cognitive recovery. The DSF, DSB, and TMT-A & B scores demonstrated a trend for improvement post-intervention, although these changes were not statistically significant. Meanwhile, the K-CES-D, EQ-5D-5L, and SES scores showed no improvement.
Usability feedback revealed critical insights into the practical aspects of the Zenicog^®^ program. Participants identified several areas for improvement, including error correction, font size, pronunciation clarity, and physical ease of use. Notably, participants with physical impairments and those less familiar with digital technology faced more significant challenges, highlighting the need for more accessible and adaptable system designs. Despite these issues, the program was generally well-received, as reflected in the usability scores, which suggests a positive user experience overall. Subgroup analyses provided further granularity, showing that usability experiences varied significantly across different demographics. These findings emphasized the importance of considering diverse user needs in the design of cognitive rehabilitation technologies to ensure broad efficacy and satisfaction.
The usability survey (Likert 1–5) (Table 4) revealed that the section on tolerance for error received the lowest mean score of 3.00 [2.38, 3.75]. When analyzed by sex (Table 5), this tendency was more pronounced in female participants, which may be attributed to sex-related predispositions. Additionally, patients with higher cognitive abilities reported issues with abrupt increases in task difficulty during program execution. This feedback will be relayed to the software developers for improvement of future versions, which should be re-evaluated in future studies. When subgrouped by educational level (Table 5), participants with fewer than 12 years of education scored the lowest in the equitability and flexibility category, with a mean score of 3.17 [2.67, 3.84]. This can be interpreted as a result of differences in digital literacy based on educational attainment. To address this issue, future studies should implement more intensive monitoring for participants with lower educational levels.
Telerehabilitation is recognized as a promising approach as compared to traditional rehabilitation methods, offering benefits such as cost reduction, enhanced accessibility, increased patient independence, and resolution of transportation and logistical issues [16]. Caregivers often face significant stress and burden during rehabilitation phases, particularly when supporting patients with cognitive impairments who have limited access to adequate healthcare services, which further exacerbate caregiver stress. Therapeutic interventions have been identified as powerful stress-reduction tools in these situations [17]. CCR is a valuable tool in this context, allowing patients to engage in repeated learning tasks independently while receiving clear feedback through automated result analysis [18]. Additionally, virtual connections enable continuous communication with therapists, supporting the maintenance of social interactions. Research also indicated that telerehabilitation can achieve cognitive improvement outcomes comparable to those achieved with face-to-face therapy [3]. Recent studies have demonstrated the cost-effectiveness and convenience of telerehabilitation, emphasizing its potential as a superior solution in terms of accessibility and efficiency [19].
Zenicog^®^ is a computerized training program developed for cognitive and language rehabilitation, with several distinctive features compared to conventional platforms.
First, whereas many existing programs are limited to specific cognitive domains and provide insufficient coverage of language, Zenicog^®^ offers an integrative program encompassing attention, memory, executive function, and four core language domains (auditory comprehension, speaking, reading, writing). This broad scope enhances its applicability across diverse patient populations with heterogeneous impairments in clinical practice.
Second, Zenicog^®^ is delivered via a tablet PC with a cloud-based monitoring system, improving accessibility, convenience, and self-directed use.
Third, it employs an AI-driven adaptive difficulty adjustment system that continuously analyzes user performance to dynamically tailor task difficulty, thereby enhancing engagement and adherence.
In addition, Zenicog^®^ supports diverse input modalities (voice, handwriting, touch) and automated scoring, enabling multidimensional feedback beyond accuracy alone.
Limitations include the initial cost of tablet PCs and potential barriers related to age or digital literacy, highlighting the need for infrastructure and user training. Some participants also reported minor issues such as eye strain or reduced touchscreen responsiveness, though these did not significantly affect adherence.
This approach enables standardized, home-based CCR, making Zenicog^®^ an effective digital healthcare solution for telerehabilitation. Furthermore, the program ensures the highest levels of data integrity and confidentiality, meeting medical device cybersecurity standards.
The study’s design, i.e., a pre–post comparison without a control group, although useful for preliminary assessment, introduces limitations in interpreting the efficacy solely attributed to the intervention. The absence of a control group means that improvements could be influenced by external factors, including natural recovery over time or participant motivation, rather than by intervention alone.
Future studies should include larger sample sizes and a randomized control group to validate these findings further. Additionally, long-term follow-up would be valuable in assessing whether the impact of the intervention on cognitive function and activities of daily living is sustained. It would also be beneficial to explore the integration of more adaptive features and personalized content to meet the specific needs and preferences of a wider range of users.
Use of an AI-guided, mobile, cognitive telerehabilitation program is feasible and is potentially beneficial in improving cognitive function for patients with post-stroke cognitive impairment or older adults with MCI. Special consideration should be given to those who are less familiar with electronic devices to improve its usability.
Overall, this study contributes to the evolving field of telerehabilitation by demonstrating the feasibility and potential benefits of an AI-guided approach for cognitive enhancement in individuals with cognitive impairments. With further refinement and research, such programs could become a vital component of rehabilitation approaches for post-stroke and MCI patients, offering effective, accessible, and user-friendly options for cognitive recovery and improvement in daily functioning.