This systematic review and meta-analysis were preregistered on PROSPERO (CRD42023492123) and were conducted according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines for transparent and comprehensive reporting of methodology and results [34]. The PRISMA 2020 checklist and the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension) checklist are provided in Multimedia Appendices 1 and 2, respectively. Any deviations from the preregistered protocol are outlined and explained in Multimedia Appendix 3.

Searches were first performed in PubMed (via NCBI), Ovid Embase (via Ovid), Web of Science (via Clarivate), CINAHL (via EBSCOhost), Cochrane Central Register of Controlled Trials (via Wiley), and APA PsycINFO (via EBSCOhost) from inception to November 21, 2023. Updated database searches were performed in the above databases via the same platforms from November 21, 2023, to January 5, 2025. Final database searches were conducted on November 22, 2025, for articles published in 2025, and email alerts were maintained to capture newly published studies until final data analysis. No study registries, websites, or other non–peer-reviewed online resources were systematically searched to ensure the reliability of the included data. The reference lists of included articles and other relevant systematic reviews were examined. Authors of potentially eligible studies were contacted via email to inquire about additional data or unpublished results.

Three groups of search terms were used: terms related to PD, terms related to DHIs (eg, digital, technology, telemedicine, internet, robotics, virtual reality, and computers), and terms related to RCTs. Complete search strategies and search limits are detailed in Multimedia Appendix 4. Search results were imported into EndNote 20 (Clarivate), and duplicates were removed using the software’s automatic deduplication feature, followed by manual review. The titles and abstracts of the involved studies were first screened, and then the full texts of potentially eligible studies were reviewed according to the inclusion criteria.

The inclusion criteria were developed based on the PICOS (Population, Intervention, Comparison, Outcomes, and Study Type) framework. First, for population, we included studies involving patients with a confirmed diagnosis of PD. Second, for intervention, interventions provided through the computer, smartphone, tablet, virtual reality, wearable sensor, robot, or any other digital technology for physical rehabilitation, care aid, or cognitive training were defined as DHIs in this review. DHIs were eligible regardless of setting (eg, used in medical institution vs at home), mode of interaction and supervision, intervention content, or intervention duration. Although digital technology also contributes to the application of deep brain stimulation, transcranial direct current stimulation, and repetitive transcranial magnetic stimulation, these interventions with specific therapeutic aims were not in the scope of this review. DHIs solely used for monitoring parameters related to the disease symptoms were not included. Third, for comparison, comparators were categorized into passive control and active control. Passive control refers to waitlist control or no intervention. Active control included traditional, nontechnological caregiving methods, such as conventional physical rehabilitation, standard care aid, or drug treatments. Studies comparing the effectiveness of different modes of DHIs were excluded. Fourth, for outcome, we considered any outcomes related to motor symptoms, nonmotor symptoms, and quality of life in patients with PD as eligible to provide a comprehensive summary. Studies that only reported the reach, fidelity, or feasibility of the DHIs in PD were also included. Fifth, for study type, only RCT studies were included in this review. Studies with fewer than 5 participants in each group were excluded from the quantitative meta-analysis to ensure the robustness of the results.

The entire process of study selection was performed by 2 researchers independently, and disagreements were resolved with the involvement of a third researcher. Interrater reliability was assessed using Cohen κ [35], which was 0.831 for full-text review, indicating a high agreement between the 2 researchers.

One reviewer extracted data using a comprehensive data extraction form: publication and author details (title, year, first author’s name, and country), study details (study design, study setting, sample size, dropout, and follow-up duration), participants’ characteristics (age, sex, and disease severity), intervention information (type of intervention, purpose of the intervention, intervention duration, and supervision mode), comparison information (type of comparator), and outcomes (scales used for assessing outcomes, reach, fidelity, and feasibility of interventions). The extracted data were checked by a second reviewer. Discrepancies in data extraction between the 2 researchers were resolved through discussion with a third reviewer.

DHIs were classified into three categories: (1) digital databases, (2) online classes, and (3) technology-based rehabilitation devices, based on the interaction modalities influenced by the technical implementation and the level of supervision inherent to the intervention, following the classification framework recommended by a previous review [36].

Digital databases are platforms that store health- or disease-related information in the form of videos or textual materials and offer self-paced access to predesigned resources, characterized by asynchronous interaction and a minimal level of supervision. Online classes involve health management and telerehabilitation sessions delivered in real time through telephone or video conferencing, offering synchronous interaction and direct supervision by professionals remotely and allowing for immediate feedback and guidance. Technology-based rehabilitation devices encompass a broad range of technologies, including exergaming, virtual reality, robotics, and mobile apps, which support physical rehabilitation or cognitive training. These devices involve close interactions with users, who respond to guidance and instructions provided by the DHIs, which in turn provide cues and feedback based on the users’ actions. The level of supervision in these interventions varies depending on the technical features of the devices (eg, sensor integration) and the study design (eg, use under professional supervision).

Additionally, DHIs were categorized into physical rehabilitation, cognitive training, and care aid according to the intervention purpose.

The primary outcomes of the review were changes in symptom severity, including motor symptoms, cognitive function, psychiatric symptoms, overall nonmotor symptoms, and quality of life from baseline to postintervention and last follow-up. The scales used for assessing the severity of each symptom and quality of life depended on the original study. When one study utilized multiple assessment scales for one type of syndrome (eg, Parkinson’s Disease Questionnaire-8 and Parkinson’s Disease Questionnaire-39 for quality of life), the most frequently used one among all included studies was selected for analysis to reduce potential heterogeneity between studies. The assessment scales for each outcome used in each study are summarized in Multimedia Appendix 5.

The secondary outcomes included percentage reach, fidelity, dropout rate, and feasibility. According to the Medical Research Council process evaluation framework and Proctor implementation outcomes, we used the more common term reach to replace penetration, which is defined as the integration of a practice within a service setting and its subsystems [37,38]; therefore, we calculated the proportion of people who come into contact with the intervention. Fidelity was defined as the degree to which an intervention was implemented as prescribed in the original protocol or as intended by the program developers [38], and we extracted the fulfillment of the intervention. Feasibility was defined as the extent to which a new treatment or innovation can be successfully used or carried out within a given agency or setting [38] and can be measured by retention, acceptability, adherence, satisfaction, preference, safety, and cost-effectiveness [39]. Feasibility was summarized according to the definition of each study, and the aforementioned items were extracted even if the study did not discuss feasibility.

The revised Cochrane Risk of Bias 2 tool [40] was used to assess study quality across the following domains: randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Studies were categorized as low risk, some concerns, or high risk. We used the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) criteria [41] to assess the certainty of the overall evidence, which incorporates 5 key considerations: study limitations, inconsistency of effects, indirectness, imprecision, and publication bias. Any deviation across these 5 domains resulted in downgrading of the quality of evidence. The overall certainty of evidence was classified as high, moderate, low, or very low. All quality assessments were independently conducted by 2 reviewers, and any discrepancies were resolved through discussion.

Pairwise meta-analyses were performed to evaluate the effect of DHIs on the primary outcomes when more than 5 effect estimates were available in the analysis. The means and SDs of within-group differences from baseline to postintervention and follow-up assessments were used. If means and SDs were not reported in a study, the corresponding author was contacted via email to obtain the data. Otherwise, means and SDs were calculated based on available data using recommended formulas [42,43].

Standardized mean differences (SMDs; Hedges g) at postintervention and the last follow-up assessment were used as the effect measure, allowing comparison of data from different assessment scales. In studies with more than 2 intervention groups, the intervention group demonstrating the most favorable effect was selected for the main analysis as it represented the most well-designed DHIs. In studies featuring multiple control groups, the most active control group was chosen to adopt a more conservative analytical approach.

Considering that different assessment scales may reflect symptom severity in opposite directions, data from the original studies for the same outcome were standardized based on the most commonly used scale. For instance, while higher scores on the Unified Parkinson’s Disease Rating Scale Part 3 indicate more severe motor symptoms, higher scores on the 6-Minute Walk Test suggest milder symptoms. Therefore, the results of the 6-Minute Walk Test were inverted (multiplied by –1) to align with the Unified Parkinson’s Disease Rating Scale Part 3 for the analysis.

Due to the presence of potential heterogeneity, the random-effects model and the DerSimonian-Laird estimator were applied to calculate pooled estimates across studies. The Hartung-Knapp-Sidik-Jonkman method was applied to increase the robustness of the results [44]. Forest plots were generated to visually display the effect size and 95% CIs for individual studies and the overall pooled estimate. To present the expected range of true effects in similar studies, the 95% prediction intervals (PIs) were also calculated and displayed in the forest plot [45]. When the number of studies was fewer than 10, the method proposed by Nagashima et al [46] was applied to adjust the 95% PI. The 95% CI reflects the precision of the estimated overall average effect, whereas the 95% PI estimates the distribution of effects within which the true effect size is expected to fall in future similar studies. These 2 intervals complement each other to provide comprehensive information for result interpretation.

Subgroup analyses were conducted based on the intervention type (technology-based rehabilitation devices, online classes, and digital databases), considering the different inherent characteristics of various DHIs. Pairwise comparisons between subgroups were performed using random-effects meta-analysis with Hartung-Knapp adjustment and were limited to subgroups with 3 or more studies. Funnel plots and Egger regression intercept were used to assess small-study effects, which may indicate the potential publication bias. A sensitivity analysis was conducted to evaluate whether the overall results were statistically and significantly affected by 1 individual study using the leave-one-out analysis method. In addition, another sensitivity analysis restricting the synthesis to studies with a low risk of bias was performed for outcomes for which at least 10 such studies were available. All meta-analyses were conducted using R (version 4.2.3; R Foundation for Statistical Computing).

The presence of heterogeneity among studies was assessed using Cochran Q test and the I² statistic, whereas 95% PI, τ, and τ² were calculated to evaluate the magnitude of heterogeneity [45]. By estimating the expected range of true effects for future studies, the 95% PI directly evaluates the variability of the intervention effect across diverse settings [47]. τ represents the estimated SD of the between-study effects, and τ² denotes the variance, which quantifies the spread of the true study effects around the mean [48].

Because substantial heterogeneity was anticipated to result from various types of DHIs in the analysis, univariate meta-regression analyses were performed using the mixed-effects model with the DerSimonian-Laird estimator for between-study variance and the Hartung-Knapp-Sidik-Jonkman method for SE adjustment to identify potential moderators that could account for heterogeneity. These potential moderators included individual-level variables (mean age, sex, and disease severity) and study-level variables (publication years, country income level, intervention type, intervention purpose, intervention setting, type of control, intervention duration, and supervision mode). In addition, bubble plots were provided as visual representation of the meta-regressions.

The search initially yielded 30,918 related study records from 6 databases in November 2023, among which 12,909 duplicate records were first removed during screening. The titles and abstracts of the remaining 18,009 records were screened, and the full texts of 284 studies were further assessed. A total of 90 studies from 91 articles evaluating the effect of DHI in patients with PD were included.

A supplementary search was conducted on January 5, 2025, and 4179 new publications were identified using the same keywords and databases as the initial search. After removing duplicates, a total of 3019 unique records were screened based on title and abstract, and the full texts of 66 potentially eligible studies were then reviewed. A total of 11 studies meeting the inclusion criteria were included.

The final search was performed on November 22, 2025, to ensure the inclusion of the most recent evidence and identified 3980 new publications. Following deduplication, a total of 2481 unique records were screened by title and abstract. Subsequently, a total of 44 studies underwent full-text review, resulting in the inclusion of 11 studies that met the eligibility criteria. During the full-text screening process, no studies were identified that appeared to meet the inclusion criteria but were ultimately excluded; all exclusions were straightforward decisions based on eligibility criteria.

Ultimately, a total of 112 unique studies from 115 articles [27,30,31,49-158] were eligible for the systematic review, and 97 unique studies from 98 articles were included in the quantitative meta-analysis. Of these 112 studies, 73 evaluated motor symptoms, 8 evaluated overall nonmotor symptoms, 26 evaluated cognitive function, 30 evaluated psychiatric symptoms, and 41 evaluated quality of life. The PRISMA flow diagram of the study search and selection is presented in Figure 1.

A total of 112 included studies comprising 5594 participants were identified, among which 84 studies (4658 participants) were performed in high-income countries, whereas only 28 studies (936 participants) were performed in low- and middle-income countries (LMICs). According to the predetermined classification criteria, the DHIs used in 91 studies were categorized as technology-based rehabilitation devices, 17 as online classes, and 8 as digital databases. The intervention purposes of 74 studies were physical rehabilitation, 28 were cognitive training, and 17 were care aid.

Active comparators (eg, traditional and nontechnological caregiving methods) were used in 91 trials, whereas others used passive comparators (eg, waitlist control or no intervention). The study setting of 51 studies was at home, and that of 55 studies was in medical institutions. Participants in the study by Halpern et al [88] first received intervention in the clinic for 9 sessions and then at home.

Regarding delivery mode, most interventions were delivered through computers and virtual reality devices; other modes included wearable sensors, robots, tablets, smartphones, and others. Most interventions lasted for at least 4 weeks, whereas 4 did not [51,54,92,115]. Among the 39 (34.8%) studies that reported follow-up assessments, the follow-up duration ranged from 4 to 24 weeks in 38 studies, except for the study by Yang et al [143], which implemented a shorter 2-week follow-up period.

A total of 97 of 112 studies were included in the meta-analysis, with the remaining 15 retained in the systematic review because of unavailable outcome data. A comprehensive summary of the characteristics of the included studies is presented in Multimedia Appendix 6.

The methodological limitations assessed by the Cochrane Risk of Bias 2 tool revealed that ≥20% of the included studies exhibited some concerns or high risk in the following domains (Figure 2): randomization process (postintervention: 53/180, 29% and follow-up: 11/55, 20%), deviations from intended interventions (postintervention: 95/180, 53% and follow-up: 19/55, 34%), missing outcome data (postintervention: 46/180, 26% and follow-up: 13/55, 24%), and selection of reported results (postintervention: 96/180, 53% and follow-up: 33/55, 60%). The main sources of bias were inadequate reporting of randomization methods (eg, lack of explicit allocation concealment) and failure to blind participants or caregivers in trials. Regarding motor symptoms, 12 studies were assessed as having a low risk of bias, 43 studies as having some concerns, and 18 studies as having a high risk of bias. Regarding cognitive function, 6 studies were assessed as having a low risk of bias, 16 studies as having some concerns, and 4 studies as having a high risk of bias. Regarding psychiatric symptoms, 4 studies were assessed as having a low risk of bias, 15 studies as having some concerns, and 10 studies as having a high risk of bias. Regarding overall nonmotor symptoms, 5 studies were assessed as having some concerns and 3 as having a high risk of bias. Regarding quality of life, 5 studies were assessed as having a low risk of bias, 22 studies as having some concerns, and 14 studies as having a high risk of bias (Multimedia Appendix 7).

Small-study effect was assessed via visual inspection of a funnel plot and Egger test (Multimedia Appendix 10). The Egger test indicated a significant small-study effect in the analysis of the effect of DHIs on motor (intercept=–2.07, 95% CI –3.69 to –0.45; P=.02) and cognitive function (intercept=2.54, 95% CI –1.12 to 3.97; P=.002) of patients with PD. Further, trim-and-fill analysis (Multimedia Appendix 10) was performed, and we observed that the effects were no longer significant for both motor symptoms (SMD=–0.04, 95% CI –0.29 to 0.21) and cognitive function (SMD=0.16, 95% CI –0.14 to 0.47).

To explore potential sources of significant heterogeneity, we performed univariate meta-regression analyses examining 3 individual-level and 8 study-level variables (Multimedia Appendix 11), and we also present bubble plots to visualize the relationship involving the continuous variables (Multimedia Appendix 12). We conducted multivariable meta-regressions using the variables mentioned above as predictors of between-study variance when 2 or more variables accounted for R². In the analysis of cognitive functions, 2 of these study variables accounted for R²: mean age of participants (model Q statistic [QM]=0.08; P=.78; R²=0.49%) and intervention setting (QM=1.16; P=.29; R²=8.88%), with a total R²=1.38% (P=.53). There are 4 variables associated with the improvement of psychiatric symptoms with a total R²=9.83% (P=.35): the percentage of female participants (QM=2.73; P=.11; R²=7.42%), publication year (QM=3.39; P=.08; R²=4.93%), country income level (QM=0.38; P=.54; R²=2.64%), and intervention purpose (QM=1.44; P=.25; R²=11.19%). As for motor symptoms, mean age of participants (QM=4.22; P=.04; R²=2.96%) and supervision mode (QM=1.50; P=.22; R²=3.96%) may explain part of the heterogeneity, with a total R²=9.42% (P=.04). Supervision mode (QM=1.41; P=.25; R²=10.38%) and mean age (QM=0.90; P=.35; R²=.64%) moderated the effect of DHIs on quality of life with a total R² of 14.56% (P=.14). These meta-regression results should be interpreted as exploratory. The low R² values indicate that the factors account for only a small proportion of the total heterogeneity. It reminds us that the variability in intervention effects is likely due to a complex combination of clinical and methodological factors not fully captured here. Nevertheless, these findings generate valuable hypotheses for future research.

Most studies reported their eligible population and participant enrollment (Multimedia Appendix 13). A total of 94 studies reported on intervention reach with a median reach of 37.5% (range 4.5%-94.7%), and all included studies reported the randomly assigned populations, with a median of 50% (range 30.6%-100%) randomly assigned to the intervention groups. A total of 38 studies reported intervention fidelity, revealing a high degree of fidelity across a diverse range, of which 34 reported percentage fidelity with a median of 95.7% (range 26.7%-100%), and another 4 studies provided relevant information regarding fidelity. Dropout rates were reported in 105 studies, with a median of 9.1% (range 0%-61.1%).

A total of 32 studies reported that the interventions were feasible to deliver with high satisfaction, adoption, preference for interventions, high interest, high retention rates, convenience, high acceptance, reliability, safety, comfort, flexibility, high adherence, cost-effectiveness, user-friendliness, high recruitment, and high fidelity. Only 5 of these reports were from LMICs. A total of 35 studies did not mention feasibility directly, but they reported a high rate of usability, no serious adverse events, favorable satisfaction ratings, willingness to recommend to others, low cost, high participant interactivity, and low dropout rate. The remaining 44 studies did not report any data on feasibility.