This study is a systematic review and network meta-analysis. The protocol was registered on PROSPERO (CRD420251006601) and adheres to the guidelines outlined in the Cochrane Handbook for Systematic Reviews of Interventions. The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines were rigorously followed to ensure methodological transparency and reporting quality (see the PRISMA checklist in Checklist 1).

In total, 6 databases—PubMed, Web of Science, Cochrane Library, Scopus, Embase, and CNKI—were systematically searched from their inception to January 2025. Two authors independently developed the search strategy to ensure methodological rigor. The search query was adapted for each database to account for differences in indexing systems, syntax, and terminology. For international databases (PubMed, Web of Science, Cochrane, Scopus, and Embase), we combined MeSH (Medical Subject Headings) and Emtree terms (eg, stroke, cognitive dysfunction, VR, and robotics) with free-text keywords (eg, PSCI, CACT, and RAT). For the CNKI database, the search strategy was translated into Chinese and optimized using discipline-specific terms (eg, 卒中后认知障碍, 虚拟现实, 计算机辅助认知训练) to align with its indexing framework. Boolean operators and proximity filters were adjusted per database requirements. No language or publication year restrictions were applied. The full search syntax for all databases, including adaptations for CNKI, is provided in Multimedia Appendix 1.

Two authors independently screened and selected studies based on predefined inclusion and exclusion criteria. Discrepancies were resolved through discussion or consultation with a third reviewer. Detailed records of study selection, including methodological quality assessment and data extraction, were maintained throughout the process.

Inclusion criteria were based on the PICOS (Population, Intervention, Control, Outcome, and Study Design) program. These are summarized in Table 1.

Exclusion criteria regarding population, intervention, control, outcome, and study design are shown in Table 2.

The evidence quality for each outcome was systematically evaluated using the Evidence quality was assessed using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) framework. If significant deficiencies were identified in domains such as risk of bias, heterogeneity, indirectness, imprecision, or publication bias, the evidence quality was downgraded from high to lower levels (eg, moderate, low, or very low) accordingly.

A total of 2128 articles were retrieved from six databases: PubMed (n=226), Web of Science (n=411), Cochrane Library (n=374), Scopus (n=161), Embase (n=580), and CNKI (n=259). After removing 682 duplicates, 1329 articles underwent title or abstract screening, with 1163 excluded for irrelevance. Full-text review of 166 articles yielded 27 [21,22] eligible RCTs (Figure 1).

A total of 27 studies with 1517 patients were included in the article, involving 5 different interventions, including VR technology, computer-assisted therapy, RAT, cognitive training, and conventional rehabilitation. There were 12 studies involving VR technology, 13 studies involving computer-assisted therapy, 4 studies involving RAT, and 8 studies involving cognitive training. The main characteristics of the included studies are shown in the table (Table 3).

Twenty-seven RCTs were included in the analysis. Based on the revised Cochrane Risk of Bias tool, 10 studies were rated as high risk, 6 as low risk, and the remaining 11 as moderate risk (some concerns). Analysis revealed that some studies exhibited deficiencies in blinding and allocation concealment, which may introduce bias into the outcomes. However, outcome measurement and reporting generally demonstrated low risk of bias across all studies. A detailed risk of bias assessment is presented in Figure 2 [21-47].

After 50,000 iterations, the convergence of the network meta-analysis models was validated as follows.

On the MoCA, CACT and VR demonstrated stable trace plots and unimodal density distributions compared to conventional therapy (eg, conventional rehabilitation, occupational therapy, and drug therapy), indicating robust convergence. RAT showed minor fluctuations in trace plots and slightly wider density distributions, suggesting potential instability. However, subsequent validation confirmed model stability. The heterogeneity parameter (I²) exhibited a right-skewed density distribution with a peak near 0.3, indicating moderate consistency across studies.

On the MMSE, RAT achieved excellent convergence, with highly stable trace plots and sharply peaked density distributions. CCT versus conventional therapy comparisons revealed mild autocorrelation in trace plots and broader density distributions, likely due to small sample biases. Heterogeneity (I²) displayed a density peak near 0.5, suggesting moderate heterogeneity attributable to protocol variations or population differences (Multimedia Appendix 2).

To understand which aspects of digital interventions may influence the effectiveness of improving cognitive function after stroke and which influences contributed to the heterogeneity of the study, a series of subgroup analyses were conducted. See Multimedia Appendix 4.

To assess the robustness of the findings, we conducted a leave-one-out sensitivity analysis by applying R to exclude each of the included RCTs in turn. The results showed significant differences in the effect of different studies on overall heterogeneity. By excluding the study of XU 2023, we found that the heterogeneity decreased from 76.65% to 18.94%, suggesting that this study might be the main source of heterogeneity. The reason for the high heterogeneity may be that the study in XU 2023 applied language training such as the Schuell Stimulation Method, which includes overall language rehabilitation, auditory stimulation training, daily communication training, speech production training, and psychological rehabilitation. This training may not only improve the patient’s language skills, but also have a positive impact on cognitive functioning (Multimedia Appendix 6).

Node-splitting analysis revealed no significant inconsistency between direct and indirect evidence across the network (P>.05 for all comparisons), confirming the stability of the consistency model. Refer to Multimedia Appendix 7.

Funnel plots for MoCA and MMSE scores exhibited symmetrical distributions, suggesting no substantial publication bias. This was further supported by Egger test results (P=.99 for the MoCA; P=.29 for the MMSE). Specific data are shown in Figures 8 and 9.

The quality of evidence was evaluated by applying the GRADE method, and the quality of evidence was all rated as low. The main reasons for this were the presence of biased variability, imprecision, and incoherence within the studies. Study bias was mainly due to the significant effect of high risk of bias studies on network estimates in all comparisons. Imprecision was due to insufficient data to demonstrate a conclusive effect and the fact that most comparisons were outside the 95% CI equivalence region.

Poststroke cognitive impairment is a complex condition of impaired cognitive functioning that occurs after stroke [48]. It usually manifests as inattention, memory loss, slow reaction time, speech-language impairment, prosopagnosia, and so on [49], which causes great obstacles to the patient’s participation in daily life and social activities [50]. With the emergence of digital technologies such as computer-assisted therapy, VR, and tele-intervention, new directions have been brought to the assessment and treatment of cognitive impairment after stroke. Digital technologies promote neuroplasticity and functional reorganization by simulating real-life scenarios, delivering multimodal stimuli and high-intensity repetitive training [51]. In addition, digital technology can quantitatively record patient training data (eg, reaction time and error rate), providing an objective basis for dynamic assessment and program optimization [52], which is more standardized and scalable than traditional training methods.

To the best of our knowledge, this study is the first net meta-analysis that simultaneously evaluates VR, computer-assisted, and robot-assisted, and includes 27 studies involving 1517 patients, aiming to assess the impact of different digital therapies on patients with poststroke cognitive dysfunction and to rank them appropriately according to their efficacy.

The results showed that for the five digital therapies mentioned above, CACT had the highest SUCRA ranking (46/50, 91.53%) in terms of MoCA improvement, suggesting that it may be the best digital therapy for enhancing executive function, possibly because it can target activation of the prefrontal network. Xiao et al [53] showed that CACT was superior to VR in MoCA scores, with significant improvement in cognitive functions, similar to the results of this study. However, CACT therapy did not demonstrate a significant advantage in MMSE scores. The present study included more RCTs, which demonstrated that CACT has a nonnegligible effect on MMSE scores through direct and indirect evidence. The main reasons why CACT can effectively improve cognitive function in stroke patients are explained as follows: CACT relies on multimedia technology and stimulates neurons in the brain through repetitive and targeted cognitive training tasks, promoting neuroplasticity and neural network reorganization [54]. It can enhance patients’ concentration on cognitive training through dynamic visual effects and different interactive tasks, and this multisensory stimulation training can activate multiple brain regions such as prefrontal and parietal lobes [55], improving patients’ attention and executive ability. A study by Lin et al [56] found that CACT significantly improved patients’ hippocampal functional linkages with the frontal and left parietal lobes through resting-state functional magnetic resonance imaging, and that improved hippocampal and frontal lobe functioning facilitated the recovery of patients’ short-term memory and executive functioning. Bernini et al [57] found significant improvements in the overall cognitive, executive, and attentional aspects of patients who underwent CACT by comparing them with traditional paper-and-pencil training. These studies once again demonstrate the reliability of the results of the present meta-analysis. Secondly, through the application of computers, the difficulty of the task can be personalized and real-time feedback can be provided [58]. By pinpointing the areas of cognitive deficits and quantitatively recording the training data of the patients, the efficiency of the training can be effectively improved, which can help to formulate a more effective assessment and intervention plan [59].

The results showed that among the five therapies mentioned above, RAT was ranked first in SUCRA in terms of MMSE scores, indicating that robotic-assisted therapy has the best efficacy in improving MMSE scores in poststroke cognitive impairment. Robotics technology can be used to capture real-time motor imagery or visually evoked Electroencephalography signals [60], translate the patient’s active cognitive intent into action commands from the robot, and have the robot assist the patient in performing the corresponding activities. This training mode can form a closed-loop stimulation of “intention-action-feedback” [61], through which the training of closed-loop stimulation can activate key brain regions such as the primary motor cortex, hippocampus, and prefrontal lobe of the patient, and promote the patient’s memory and overall cognitive function improvement [62].

The study included 2 different outcome metrics, which were analyzed and found to yield different conclusions. On the MoCA, CACT had the best intervention effect, which may be due to the fact that the MoCA covers both visuospatial and abstract thinking dimensions and focuses more on the assessment of executive ability, whereas CACT improves executive functioning mainly through activation of the prefrontal network [63]; and thus, the MoCA is more sensitive to the therapeutic effect of CACT. The MMSE prioritizes global cognition (eg, orienting and recalling), whereas the motor-cognitive dual-task paradigm of RAT better addresses this through sensorimotor integration and hippocampal activation. Studies by Jia et al [64] and Agogo et al [65] showed that the MMSE has a “ceiling” effect on executive and recall functions, whereas the MoCA is more sensitive to the detection of executive abilities. This justifies the accuracy of the results of this study.

The effectiveness of VR for cognitive function has been controversial. Some studies [66,67] have pointed out that VR is currently focused on the improvement of motor function, and there are fewer studies on cognitive function, which makes it difficult to provide a reliable evidence-based basis. A study by Wiley et al [68] showed that the improvement of cognitive function by VR was not significantly different from that of the control group, and it only improved the patients’ initiative and sense of experience in rehabilitation. The study by Zhang et al [69] supports the viewpoint of Wiley et al [68]. This study found that VR therapy had significant efficacy on the MoCA but little efficacy in the MMSE.

The differential efficacy of VR across the MoCA and MMSE outcomes likely arises from both mechanistic and methodological factors. Mechanistically, the MoCA emphasizes executive function (eg, working memory, cognitive flexibility, and visuospatial reasoning), which aligns with VR’s capacity to deliver task-specific challenges requiring real-time decision-making and adaptive problem-solving [67]. For example, VR systems like Reh@City integrate dual-task paradigms (eg, navigating virtual environments while solving arithmetic problems), which selectively activate the dorsolateral prefrontal cortex and posterior parietal regions—key neural substrates for executive function [70]. These targeted activations may explain VR’s superior performance in MoCA scores. Conversely, the MMSE prioritizes global cognitive domains such as orientation, memory recall, and language comprehension—functions less directly engaged by current VR protocols. Most VR interventions in this analysis focused on motor-cognitive integration (eg, reaching tasks in virtual environments), which may inadequately stimulate hippocampal and medial temporal networks critical for MMSE-assessed domains.

Methodologically, heterogeneity in VR protocols further complicates interpretation. Studies demonstrating VR benefits for the MoCA; for example, Faria et al [24] and Rogers et al [26] used adaptive difficulty systems and multisensory feedback, whereas trials with null MMSE outcomes, for instance, those of Choi et al [40] and Oh et al [28] used simpler VR tasks lacking cognitive progression or sensory feedback. In addition, the MMSE’s ceiling effects in mild cognitive impairment populations [71] and its insensitivity to executive improvements may underestimate VR’s impact on higher-order cognition. Second, cognition consists of several aspects, such as memory, executive ability, etc [72]. Only by corresponding to specific areas can a certain cognitive function be effectively improved. Perhaps a VR system that truly focuses on cognitive rehabilitation can change the controversial situation of VR for improving cognition.

There are significant differences in the cognitive domains assessed by the MoCA and the MMSE. The MoCA focuses on assessing higher cognitive functions such as executive functioning, attention, and visuospatial abilities, while the MMSE focuses primarily on basic cognitive functions such as orientation, memory, and language. Therefore, they have different sensitivities and specificities in assessing cognitive impairment.

In the study, it was found that CACT showed a significant effect in improving executive functioning through activation of the prefrontal network, which was reflected in the MoCA scores. However, this effect may not directly translate into significant improvements in MMSE scores, which are more focused on basic cognitive functions. Similarly, the RAT activates the motor cortex and hippocampus through a dual motor-cognitive task paradigm, thereby showing benefits in global cognitive function as assessed by the MMSE, but may not be as effective as the CACT in executive function as assessed by the MoCA. So different interventions may be effective in specific cognitive domains, but not necessarily for all cognitive impairments.

The study’s findings provide actionable insights for integrating digital therapies into clinical practice, contingent on balancing efficacy, accessibility, and patient-specific factors.

CACT is optimal for patients with executive dysfunction (prioritize MoCA-based deficits). Most CACT applications have software license fees (eg, the Reh@Com system), but have low hardware requirements and are reusable for patients with long-term interventions. Their personalized algorithms can reduce labor inputs and are cost-effective and scalable in resource-limited areas.

RAT has a high initial acquisition cost, but it can integrate motor and cognitive training and reduce multiple equipment investments. This meta-analysis showed that RAT had the highest SUCRA score (99.44%) for the MMSE, suggesting that it is efficient in improving patients’ overall cognitive function and may have long-term cost-effectiveness. Despite the high initial investment, RAT plays a critical role in long-term rehabilitation and may offset the initial cost. In the clinic, this integrated training approach increases treatment efficiency, reduces the cost of switching patients between multiple devices, and improves the rehabilitation experience and adherence. Thus, in the long term, RAT excels in cognitive rehabilitation and may be a cost-effective option.

Although VR technology has gained attention for its highly engaging and superior visuospatial training effects, it faces some challenges in clinical applications. For example, the high cost of immersive VR devices (eg, Oculus Quest 2) and associated scenario development may limit its widespread use in certain resource-limited health care organizations.

However, it is worth pointing out that the efficiency of immersive training can somewhat compensate for its costly investment. By providing a highly immersive training environment, VR technology can significantly reduce the time required for a single training session while improving the effectiveness of the training. This efficiency not only helps patients recover cognitive functions faster, but also indirectly reduces the total cost of rehabilitation and care in the long run. Thus, despite the problem of high initial costs, the benefits of this technology may far outweigh its costs in the long run, especially in cases where frequent training and long-term interventions are required.

Patient preferences also play an important role in the application of therapy. For example, younger patients or those who are more receptive to technology may prefer more interactive digital therapies such as VR or CACT, whereas older patients or those who are unfamiliar with technology may be more comfortable with traditional cognitive training or RAT. In addition, the type and severity of a patient’s cognitive impairment may also affect their acceptance of and adherence to different therapies. Therefore, digital cognitive intervention methods should consider the above factors in the process of clinical application and develop individualized programs suitable for patients.

This study provides valuable insights into the efficacy of digital therapies for PSCI, but several limitations must be acknowledged. First, the GRADE framework rated the quality of evidence as low across all comparisons, primarily due to three interrelated issues: (1) risk of bias: approximately 26% of included studies exhibited high risk of bias, particularly in blinding of participants/personnel and allocation concealment. For example, unblinded therapists may have inadvertently influenced patient engagement in VR or RAT interventions, potentially inflating effect estimates for these therapies compared to conventional rehabilitation. (2) Imprecision: wide CIs in key comparisons reflect insufficient statistical power to distinguish small but clinically meaningful differences, likely due to small sample sizes in robotic trials (n=4 studies). (3) Incoherence: Heterogeneity persisted despite sensitivity analyses, attributable to variations in intervention protocols (eg, adaptive vs fixed-difficulty VR systems) and population characteristics (eg, inclusion of both subacute and chronic stroke phases).

These limitations have important implications for clinical implementation. Although CACT and RAT demonstrated domain-specific superiority in MoCA and MMSE scores, respectively, the lower GRADE scores suggest that clinicians should view the study’s findings with caution. For example, the apparent superiority of the RAT on the MMSE may be influenced by its frequent combination with motor training (eg, the Lokomat system), which may synergistically enhance overall cognitive performance through activation of sensory-motor networks, a mechanism that has not been isolated in the current RAT trial.

In addition, this study mainly focused on the short-term treatment effect and lacked long-term follow-up data to comprehensively assess the long-term efficacy of digital therapies and the long-term prognosis of patients, precluding conclusions regarding sustained cognitive benefits or cost-effectiveness.

Given these limitations, future research should focus on the following areas. First, it is crucial to conduct long-term follow-up studies. This will help to gain insights into the long-term effects of digital therapies on patients with PSCI and assess their lasting effects in delaying cognitive decline, improving quality of life, and promoting social regression. Second, future studies should further optimize the intervention protocols, including clarifying the optimal intervention dose, frequency, and duration of different digital therapies, as well as exploring the basis for the development of individualized treatment plans. In addition, combining neuromodulation techniques or traditional rehabilitation means, such as transcranial magnetic stimulation and acupuncture, with digital therapies for synergistic application may bring more significant rehabilitation benefits to patients, which is worth exploring in depth. Finally, from the perspective of technological development, with the continuous advancement of AI, VR, and other technologies, how to ensure that digital therapies reduce costs and maintain therapeutic efficacy through continuous technological updating and optimization is also an important area of research. For example, in resource-limited settings, CACT delivered via tablet devices can provide a cost-effective cognitive rehabilitation program. And reinforcement learning algorithms can be integrated into VR-based cognitive training modules to dynamically adjust task difficulty based on real-time performance. Future research should develop smarter adaptive training systems, more immersive VR scenarios, and more accurate tools for monitoring and analyzing patient data in order to further improve the effectiveness and feasibility of digital therapies in the field of poststroke cognitive rehabilitation.

Overall, digital therapies provide scientific and personalized solutions for poststroke cognitive rehabilitation, with CACT and RAT demonstrating domain-specific advantages in executive function and global cognition, respectively. These findings align with the study’s aims to quantify relative efficacy and identify optimal strategies for specific cognitive domains. However, heterogeneity and methodological limitations in existing trials necessitate further validation through high-quality RCTs with standardized protocols, larger samples, and long-term follow-ups. Clinical translation requires balancing efficacy, safety, and cost-effectiveness to advance precision rehabilitation strategies.