Dose–response effects of robot-assisted training on poststroke functional recovery

Abstract

Background:

Robot-assisted training (RAT) is widely used in poststroke rehabilitation, but the relationship between training dose and functional recovery remains unclear. This review evaluated the effectiveness of RAT and examined dose–effect relationships.

Methods:

PubMed, Web of Science, the Cochrane Library, and Embase were searched from inception to October 2025. Randomized controlled trials of adults with stroke comparing RAT or RAT combined with conventional rehabilitation (CR) were included. Study quality was assessed using the Cochrane risk-of-bias tool, and analyses were performed using Review Manager 5.4 and Stata 16.0. Pooled effects were expressed as mean differences (MDs) or standardized MDs with 95% confidence intervals (CIs). Exploratory subgroup analyses and random-effects meta-regression were conducted across 5 training-dose parameters, and restricted cubic spline analyses assessed nonlinear dose–effect relationships. Evidence certainty was evaluated using the Grading of Recommendations Assessment, Development and Evaluation. The review was registered with International Prospective Register of Systematic Reviews (CRD420251267710).

Results:

Thirty-six randomized controlled trials involving 1941 patients were included. Compared with CR, RAT significantly improved upper limb motor function (MD = 2.89, 95% CI: 1.49–4.28) and activities of daily living (ADLs; standardized mean difference = 0.28, 95% CI: 0.13–0.43), but not balance function (MD = 1.27, 95% CI: −0.92 to 3.47). Exploratory subgroup analyses showed numerical differences across dose levels, but no between-subgroup differences were significant. None of the 5 training-dose parameters showed significant linear associations with Fugl–Meyer Assessment–Upper Extremity, Berg Balance Scale, or ADL, and restricted cubic spline analyses showed no significant nonlinear associations (all P > .05). Evidence certainty was moderate for Fugl–Meyer Assessment–Upper Extremity and ADL and low for Berg Balance Scale. Dose-related findings should be interpreted cautiously because some subgroups included few studies and actual training dose was incompletely reported.

Conclusion:

RAT may improve upper limb motor function and ADL after stroke, but has not shown a significant advantage for balance. Current evidence does not demonstrate a clear linear or nonlinear relationship between training dose and rehabilitation outcomes.

1. Introduction

About 90% of stroke survivors suffer certain functional disorders, many of which affect the performance of stroke patients in activities of daily living (ADLs).^[1] Robot-assisted training (RAT) can improve relevant functions and positively impact ADLs and quality of life.^[2] However, poststroke functional recovery is a dynamic and complex process, and a patient’s functional status may dynamically fluctuate among improvement, deterioration, and stabilization.^[3] Thus, while the overall efficacy of RAT has been demonstrated,^[4] further clarifying the differences in functional improvement under different training doses, as well as the changing patterns, is significant for achieving more precise rehabilitation interventions.

The poststroke RAT dose is both a critical issue in clinical practice and a key challenge in current research.^[5] First, differences in patient groups, stroke severity, intervention protocols, and robot types all bring challenges in determining an appropriate training dose for RAT therapy.^[4] Second, the RAT dose involves multiple parameters (intensity, frequency, treatment duration, etc), and different combinations and settings of these parameters lead to different therapeutic effects. Hence, there are no standardized definitions and quantification systems for RAT dose in the field of stroke rehabilitation.^[6,7] The existing evidence indicates no consensus has been reached on RAT dose regimens. Some research suggests that RAT has a cumulative dose effect and that high dose and high intensity are its primary reasons for functional improvement in stroke patients.^[8] However, other research shows a diminishing-returns plateau effect may exist in the efficacy of RAT, where a further increase in training duration or dose does not consistently enhance its intervention effectiveness, and a simple superimposition of training intensity and dose may not necessarily lead to better improvement outcomes.^[9] The European Stroke Organization rehabilitation guideline holds that the current evidence regarding stroke treatment dose remains insufficient, and there is no high-quality evidence supporting dose issues.^[10] This finding indicates that although the evidence for the overall efficacy of RAT continually accumulates, it has not been directly translated into clear and operable dose prescription recommendations, and the treatment dose recommendations in the guideline are largely based on low-quality evidence. As for specific dose parameters, a total intervention frequency ≤15 sessions may be associated with greater improvements in gait speed and balance function.^[11] Meanwhile, a training regimen of 45 to 60 minutes/day and ≥3 hours per week may achieve better rehabilitation outcomes.^[12] Notably, the benefits of treatment dose may be influenced by the individualization of intervention schemes. For instance, RAT with an individualized scheme demonstrates superior upper limb functional recovery over RAT without a definite scheme.^[13] Moreover, some research suggests through dose-escalation approaches that moderate to severe stroke patients can tolerate upper limb rehabilitation training at doses higher than routine therapy, though this research is focused on safety rather than efficacy.^[14] Meanwhile, other research suggests higher doses are not necessarily better, and reducing the single-session dose and increasing daily training frequency may more effectively improve the functional recovery of stroke patients.^[15] Additionally, low-dose schemes show similar functional improvements and are more cost-effective compared with high-dose schemes in some cases.^[2] However, most RAT studies report single-session duration and total training dose, but there is rare systematic reporting and in-depth explanation of key elements (e.g., effective dose, personalized intervention scheme, and dose–response relationship),^[16] which reflects a significant gap between dose standardization and clinical translation in this field. Bernal-Jiménez et al suggest that future research should focus more on RAT treatment intensity and dose and clarify how different variables influence the efficacy of RAT in stroke patients.^[17] Overall, existing studies have reported inconsistent findings regarding RAT training dose, and the potential effects of different dose regimens should therefore be interpreted cautiously.

Under such a background, this systematic review and meta-analysis aimed to comprehensively evaluate the overall efficacy of RAT on upper limb motor function, balance function, and ADL in poststroke patients. Furthermore, subgroup analyses from multiple dose dimensions were conducted to explore the dose–response relationship and reveal efficacy differences and variation patterns under different doses, thereby providing evidence-based support for developing individualized precise RAT clinical schemes.

2. Materials and methods

This study was registered on International Prospective Register of Systematic Reviews (No. CRD420251267710) and strictly adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 statement for reporting; the corresponding checklist is provided in Table S1, Supplemental Digital Content 1.

This study was based solely on data from previously published studies and did not involve the direct recruitment of human participants or identifiable individual-level data. Therefore, ethical approval and informed consent were not required.

2.1. Search strategy

English databases PubMed, Web of Science, the Cochrane Library, and Embase were searched from the inception of each database to October 2025. A systematic search strategy combining Medical Subject Headings and free-text terms was developed. The search terms primarily included “robot-assisted training,” “stroke,” “randomized controlled trial,” and relevant expanded terms. The detailed search strategy was provided in Table S2, Supplemental Digital Content 2.

2.2. Eligibility criteria

2.2.1. Inclusion criteria

The inclusion criteria were as follows: study type, randomized controlled trial (RCT); subjects, patients clinically diagnosed with stroke based on medical history, symptoms, and signs, and aged ≥18 years, without restrictions on gender, ethnicity, nationality, or disease duration; interventions, the treatment group receiving RAT or RAT combined with conventional rehabilitation (CR) training and the control group receiving CR; and outcome indices, including at least one of the following: upper limb motor function (Fugl–Meyer Assessment–Upper Extremity [FMA-UE]), balance function (Berg Balance Scale [BBS]), or ADL (Barthel Index/modified Barthel Index [BI/MBI]).

2.2.2. Exclusion criteria

The exclusion criteria were as follows: non-English literature; duplicate publications; conference abstracts, patents, dissertations, reviews, meta-analyses, or interim reports of a study; subjects without a clear clinical diagnosis or with concurrent major diseases; and unclear intervention measures or duration, incomplete outcome indicators, or data unsuitable for relevant statistical analysis.

2.3. Data extraction

Two researchers independently screened the literature according to the inclusion and exclusion criteria and extracted data. The data included study background: author, year; patient characteristics: age, sample size, mean disease duration; intervention methods: intervention type, training dose, training frequency; indicators reflecting study quality; and relevant outcome measures of interest and measurements. After data collection, the 2 researchers cross-checked the data. Any disagreement was resolved through discussion with a third researcher to reach consensus.

2.3.1. Definition of training dose parameters

We define the training dose of RAT as a comprehensive concept composed of 5 quantifiable parameters: session dose (minutes/session); weekly training frequency (sessions/week); intervention duration (weeks); weekly training volume (minutes/week); and total training dose (minutes). Notably, the lack of standardized reporting in the studies complicates quantitative comparisons. Hence, this analysis focuses on the above dose parameters and does not further examine specific training content (e.g., training type and training difficulty) within the RAT interventions.

2.3.2. Classification of training dose parameters

Because there are currently no standardized dose cutoff criteria for RAT, the subgroup thresholds in this study were determined primarily with reference to the dose categorization methods used in previous meta-analyses of robot-assisted stroke rehabilitation, together with the distributions of the dose parameters across the included studies. Therefore, these categories were mainly used to explore differences in treatment effects across different dose levels rather than to represent established clinical dose thresholds.

2.4. Quality assessment

The RCT bias risk assessment tool recommended by Cochrane Handbook 5.1.0 was adopted to evaluate bias due to random sequence generation, allocation concealment, blinding implementation, completeness of results, selective reporting, and other sources of bias in the included studies. The risk of bias was categorized as low, unclear, or high.

2.5. Data synthesis

Statistical analysis was conducted using Review Manager 5.4 (The Cochrane Collaboration) and Stata 16.0 (StataCorp LLC). The outcome measures were all continuous variables. FMA-UE and BBS were analyzed using mean difference (MD), and BI/MBI were evaluated using standardized MD (SMD). The effect sizes were all expressed as 95% confidence intervals (CIs). Heterogeneity was assessed using the χ² test and I² test. In case of P > .05 or I² < 50%, a fixed-effect model was employed; otherwise, a random-effects model was used. The robustness of the pooled results was evaluated through sensitivity analysis, while publication bias was assessed using funnel plots and Egger’s test. The quality of evidence was rated using the Grading of Recommendations Assessment, Development and Evaluation approach. Random-effects meta-regression was performed to assess the linear associations between each of the 5 training-dose parameters and treatment effects, with between-study variance estimated using restricted maximum likelihood and Knapp–Hartung adjustment. Restricted cubic spline (RCS) models were used to assess potential nonlinear dose–effect relationships, with 3 knots placed at the 10th, 50th, and 90th percentiles of the dose distribution. Weekly training frequency was not included in the RCS analysis because of its discrete and highly concentrated distribution. To account for potential confounding by time since stroke onset, adjusted random-effects meta-regression analyses were performed for each training-dose parameter, with the natural logarithm of time since stroke onset included as a covariate; studies without available time-since-stroke data were excluded from the corresponding adjusted analyses.

3. Results

3.1. Search results

A total of 1171 articles were retrieved from the 4 databases. After removing duplicates using EndNote 21 (Clarivate), 860 articles remained. After initial screening of titles and abstracts, 91 articles were sent for full-text evaluation. Through full-text reading and screening, 36 studies were finally included (Fig. 1).

Figure 1.

PRISMA flow diagram of literature screening. PRISMA = Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

3.2. Characteristics of included studies

The 36 included studies involved 1941 stroke patients, with 979 patients in the treatment group and 962 patients in the control group.^[18–53] The included patients were mainly aged 50 to 65 years. The session dose of RAT ranged from 15 to 120 minutes, the weekly training frequency was 2 to 7 sessions, the intervention duration lasted 2 to 12 weeks, the total weekly training volume ranged from 75 to 525 minutes, and the overall training dose was 200 to 3840 minutes. As for outcome measures, 20 studies reported upper limb motor function scores^{[18–22,26–29,35,36,39,40,42,43,47–50,52]}; 13 studies reported balance function scores^{[23–25,30–34,37,41,44,51,53]}; and 15 studies reported ADL scores.^{[18,24–26,30,33,34,38,40,42,45,46,48,52,53]} Detailed characteristics of the included studies are presented in Table 1.

Table 1.

Basic information of included studies.

Study	n (M/F)	Age (yr, mean ± SD)	Disease duration (mean ± SD)	Intervention types	Time/session (min)	Sessions/wk (n/wk)	Duration (wk)	Dose/wk (min/wk)	Total dose (min)	OMs
Başar et al[18]	T:12/9	69.48 ± 8.55	12.57 ± 19.76 (mo)	RAT + CR	40 (RAT)	5	4	200	800	①③
	C:10/11	66.05 ± 9.14	16.57 ± 19.68 (mo)
Bhattacharjee et al[19]	T:13/9	53.32 ± 9.93	4.14 ± 1.21 (mo)	RAT + CR	30 (RAT)	5	3	150	450	①
	C:12/10	53.23 ± 10.51	4.09 ± 1.57 (mo)
Valles et al[20]	T:7/3	44.1 ± 12.55	NR	RAT + CR	120 (RAT)	3–4	6–8	360–480	2160–3840	①
	C:6/4	64.1 ± 8.38	NR
Carpinella et al[21]	T:10/9	64.84 ± 9.61	6.86 ± 8.16 (mo)	RAT + CR	45 (RAT)	5	4	225	900	①
	C:10/9	57.92 ± 18.42	34.4 ± 70.24 (mo)
Feingold-Polak et al[22]	T:10	54.3 ± 12.7	108 ± 56 (d)	RAT + CR	45–60 (RAT)	2–3	5–7	90–180	450–1260	①
	C:8	57.3 ± 12.7	92 ± 42 (d)
Gandolfi et al[23]	T:10/6	63.87 ± 11.44	54.71 ± 36.28 (mo)	RAT	45 (RAT)	2	5	90	450	②
	C:13/3	64.37 ± 10.56	53.06 ± 41.73 (mo)
Han et al[24]	T:17/13	67.89 ± 14.96	21.56 ± 7.98 (d)	RAT + CR	30 (RAT)	5	4	150	600	②③
	C:15/11	63.2 ± 10.62	18.10 ± 9.78 (d)
Haolin et al[25]	T:22/3	58 ± 12	7 ± 4 (wk)	RAT + CR	20 (RAT)	5	2	100	200	②③
	C:20/5	59 ± 10	8 ± 8 (wk)
Hesse et al[26]	T:13/12	71.4 ± 15.5	4.5 ± 1.7 (wk)	RAT + CR	30 (RAT)	5	4	150	600	①③
	C:15/10	69.7 ± 16.6	4.5 ± 1.4 (wk)
Hsieh et al[27]	T:11/5	52.34 ± 13.20	23.56 ± 15.43 (mo)	RAT	90–105 (RAT)	5	4	450–525	1800–2100	①
	C:12/4	54.12 ± 9.98	27.81 ± 19.07 (mo)
Hsieh et al[28]	T:11/5	49.28 ± 10.90	2.56 ± 1.69 (mo)	RAT + CR	40–45 (RAT)	5	4	200–225	800–900	①
	C:7/8	52.87 ± 10.40	2.21 ± 1.11 (mo)
Hsu et al[29]	T:11/11	53.1 ± 13.9	13.7 ± 8.6 (mo)	RAT + CR	40 (RAT)	3	4	120	480	①
	C:9/12	52.6 ± 12.5	14.7 ± 13.2 (mo)
Huo et al[30]	T:10/4	57.93 ± 11.47	68.07 ± 35.67 (d)	RAT	30 (RAT)	5	4	150	600	②③
	C:11/5	55.25 ± 11.16	74.00 ± 33.29 (d)
Kang et al[31]	T:10/5	64.3 ± 4.6	168.3 ± 67.3 (mo)	RAT	30 (RAT)	NR	3	300	900	②
	C:8/7	62.9 ± 6.0	142.6 ± 59.2 (mo)
Kim et al[32]	T:8/12	56.8 ± 12.78	NR	RAT + CR	15 (RAT)	5	8	75	600	②
	C:10/10	56.75 ± 13.31	NR
Kim et al[33]	T:16/6	59.22 ± 14.0	34.18 ± 36.45 (mo)	RAT	30 (RAT)	3	8	90	720	②③
	C:11/11	64.0 ± 10.2	24.38 ± 17.99 (mo)
Kim et al[34]	T:20/5	57.7 ± 12.9	2.0 ± 2.4 (mo)	RAT + CR	30 (RAT)	5	3	150	450	②③
	C:13/10	60.4 ± 13.2	2.6 ± 3.1 (mo)
Kim et al[35]	T:9/6	65.53 ± 8.43	214.2 ± 69.52 (mo)	RAT	30 (RAT)	NR	3	300	900	①
	C:9/6	64.53 ± 7.72	210.6 ± 80.78 (mo)
McCabe et al[36]	T:10/2	NR	NR	RAT	90–105 (RAT)	5	4	450–525	1800–2100	①
	C:6/5	NR	NR
Lee et al[37]	T:7/7	61.85 ± 7.87	1486 ± 264.12 (d)	RAT	45 (RAT)	3	4	135	540	②
	C:7/5	62.25 ± 6.36	1536 ± 311.54 (d)
Li et al[38]	T:16/4	63.6 ± 11.4	67.35 ± 47.31 (d)	RAT + CR	20 (RAT)	5	4	100	400	③
	C:17/3	63.7 ± 9.4	70.55 ± 49.37 (d)
Liao et al[39]	T:6/4	55.51 ± 11.17	23.90 ± 13.39 (mo)	RAT	90–105 (RAT)	5	4	450–525	1800–2100	①
	C:7/3	54.56 ± 8.20	22.20 ± 17.47 (mo)
Lin et al[40]	T:60/22	59.37 ± 10.96	142.30 ± 162.84 (d)	RAT + CR	30 (RAT)	5	3	150	450	①③
	C:64/22	58.72 ± 12.89	158.23 ± 178.20 (d)
Lin et al[41]	T:14/6	54.1 ± 8.6	25.8 ± 26.5 (d)	RAT + CR	30 (RAT)	5	3–4	150	450–600	②
	C:15/5	56.5 ± 12.9	34.7 ± 25.5 (d)
Liu et al[42]	T:20/4	65.8 ± 9.0	3.9 ± 2.5 (mo)	RAT + CR	30 (RAT)	5	4	150	600	①③
	C:15/9	66.1 ± 6.9	3.5 ± 2.7 (mo)
Masiero et al[43]	T:10/4	65.60 ± 9.2	8.342 ± 3.2 (d)	RAT + CR	40 (RAT)	5	5	200	1000	①
	C:10/6	66.83 ± 7.9	10.23 ± 2.4 (d)
Min et al[44]	T:11/8	61.47 ± 11.15	921.52 ± 1762.0 (d)	RAT + CR	30 (RAT)	5	4	150	600	②
	C:13/6	56.36 ± 9.16	788.73 ± 999.24 (d)
Morone et al[45]	T:12	55.58 ± 13.35	NR	RAT + CR	NR	5	4	NR	NR	③
	C:12	60.17 ± 9.59	NR
Mustafaoglu et al[46]	T:10/7	51.9 ± 15.3	13.45 ± 7.52 (mo)	RAT + CR	45 (RAT)	2	6	90	540	③
	C:12/5	55.3 ± 11.2	13.62 ± 7.24 (mo)
Rodgers et al[47]	T:156/101	59.9 ± 13.5	233 ± 331.11 (d)	RAT	45 (RAT)	3	12	135	1620	①
	C:153/101	62.5 ± 12.5	242 ± 327.40 (d)
Singh et al[48]	T:12	41.1 ± 12.8	13.8 ± 9.1 (mo)	RAT	45 (RAT)	5	4	225	900	①③
	C:11	42.7 ± 9.3	10.3 ± 5.0 (mo)
Takahashi et al[49]	T:21/9	65.2 ± 10.9	47.8 ± 7.0 (d)	RAT + CR	40 (RAT)	7	6	280	1680	①
	C:18/8	64.6 ± 11.5	46.9 ± 8.1 (d)
Takebayashi et al[50]	T:33/8	60 ± 11	NR	RAT + CR	40 (RAT)	3	10	120	1200	①
	C:26/10	58 ± 11	NR
Wright et al[51]	T:14/2	59.6 ± 10.1	31 ± 19 (mo)	RAT + CR	30 (RAT)	5–6	10	150–180	1500–1800	②
	C:14/4	65.1 ± 10.1	32 ± 21 (mo)
Yuan et al[52]	T:21/3	59.0 ± 8.3	5.5 ± 5.5 (wk)	RAT + CR	30 (RAT)	6	3	180	540	①③
	C:20/3	58.9 ± 10.3	5.0 ± 3.6 (wk)
Zhang et al[53]	T:10/2	63.50 ± 8.97	46.67 ± 24.62 (d)	RAT + CR	60 (RAT)	5	4	300	1200	②③
	C:8/4	63.83 ± 8.28	52.50 ± 28.96 (d)

Values are expressed as mean ± SD. ① FMA-UE, Fugl–Meyer Assessment–Upper Extremity; ② BBS, Berg Balance Scale; ③ BI/MBI, Barthel Index/modified Barthel Index.

C = control group, CR = conventional rehabilitation, d = day, F = female, h = hour, M = male, mo = month, NR = not reported, OMs = outcome measures, RAT = robot-assisted training, SD = standard deviation, T = treatment group, wk = week.

3.3. Risk of bias in included studies

Of the 36 included studies, all studies (100%) had a low risk of bias in random sequence generation and mainly used random number tables and block randomization. Twenty studies (55.56%) used allocation concealment methods, including opaque envelopes and central randomization systems. Since RAT has obvious operational characteristics, 29 studies (80.55%) did not apply the blinding method to personnel or participants. However, 29 studies (80.55%) employed the blinding method for outcome assessors. As for the completeness of data, 34 studies (94.44%) provided complete results. The 36 studies (100%) did not selectively report results. In terms of other biases, most studies showed no significant risks, as only individual studies indicated baseline imbalance or potential conflict of interest (Figs. 2 and 3).

Figure 2.

Proportional distribution of each type of bias risk.

Figure 3.

Summary of bias risks in included studies. Green circles represent a low risk of bias, yellow circles indicate some concerns, and red circles denote a high risk of bias.

3.3.1. Upper limb motor function

A total of 20 studies were included and used the FMA-UE scale to assess patients’ upper limb motor function, with a total sample size of 1270 cases. The heterogeneity among studies was low (P = .34, I² = 9%), so a fixed-effects model was used for pooled analysis. Meta-analysis showed RAT was significantly superior to CR in improving upper limb motor function (MD = 2.89, 95% CI: 1.49–4.28, P < .001; Fig. 4).

Figure 4.

Forest plot of the effect of RAT on upper limb motor function (FMA-UE). CI = confidence interval, FMA-UE = Fugl–Meyer Assessment–Upper Extremity, IV = inverse variance, RAT = robot-assisted training, SD = standard deviation.

3.3.2. Balance function

A total of 13 studies were included and used the BBS scale to assess patients’ balance function, with a total sample size of 488 cases. The heterogeneity among studies was high (P = .02, I² = 52%), so a random-effects model was used for pooled analysis. Meta-analysis showed RAT was not significantly different from CR in improving balance function (MD = 1.27, 95% CI: −0.92 to 3.47, P = .26; Fig. 5).

Figure 5.

Forest plot of the impact of RAT on balance function (BBS). BBS = Berg Balance Scale, CI = confidence interval, IV = inverse variance, RAT = robot-assisted training, SD = standard deviation.

3.3.3. ADLs

A total of 15 studies were included and used BI or MBI to assess patients’ ADLs, with a total sample size of 727 cases. Moderate heterogeneity was observed among the studies (P = .05, I² = 41%), so a fixed-effects model was used for pooled analysis. Meta-analysis showed RAT was significantly superior to CR in improving ADL (SMD = 0.28, 95% CI: 0.13–0.43, P < .001; Fig. 6).

Figure 6.

Forest plot of the effects of RAT on ADLs (BI/MBI). ADLs = activities of daily living, BI/MBI = Barthel Index/modified Barthel Index, CI = confidence interval, IV = inverse variance, RAT = robot-assisted training, SD = standard deviation.

3.4. Subgroup analyses

Five exploratory subgroup analyses based on intervention dose parameters were conducted to compare treatment effects across different dose levels, including session dose (minutes/session), weekly training frequency (sessions/week), intervention duration (weeks), weekly training volume (minutes/week), and total training dose (minutes). The results are presented in Table 2.

Table 2.

Exploratory subgroup analyses of robot-assisted training according to different dose parameters.

Dose variable	Dose subgroup	Studies (n)	Outcome	Effect size (type, 95% CI)	Z-value	Within-subgroup P value	I2 (%)	P for subgroup differences
Session dose	≤30 min	6	FMA-UE	MD 3.46 (1.24–5.68)	3.05	<.01	26	.68
		10	BBS	MD 0.92 (−1.48 to 3.32)	0.75	.45	53	.49
		10	BI/MBI	SMD 0.20 (0.04–0.37)	2.44	.01	32	.13
	31–60 min	10	FMA-UE	MD 2.79 (0.76–4.83)	2.69	<.01	0	.68
		3	BBS	MD 3.55 (−3.54 to 10.63)	0.98	.33	52	.49
		4	BI/MBI	SMD 0.51 (0.15–0.87)	2.75	<.01	32	.13
	>60 min	4	FMA-UE	MD 1.52 (−2.29 to 5.33)	0.78	.43	35	.68
Weekly training frequency	≤3/wk	4	FMA-UE	MD 2.60 (0.21–4.98)	2.14	.03	48	.89
		3	BBS	MD −0.42 (−3.89 to 3.05)	0.24	.81	0	.34
		2	BI/MBI	SMD 0.30 (−1.35 to 1.95)	0.36	.72	92	.76
	4–5/wk	13	FMA-UE	MD 3.10 (1.27–4.93)	3.32	<.001	4	.89
		8	BBS	MD 2.96 (−0.21 to 6.12)	1.83	.07	50	.34
		12	BI/MBI	SMD 0.28 (0.12–0.44)	3.39	<.001	0	.76
	≥6/wk	2	FMA-UE	MD 3.95 (−1.82 to 9.71)	1.34	.18	35	.89
		1	BBS	MD 2.80 (−2.99 to 8.59)	0.95	.34	–	.34
		1	BI/MBI	SMD 0.51 (−0.07 to 1.09)	1.72	.09	–	.76
Intervention duration	≤4 wk	14	FMA-UE	MD 2.92 (1.18–4.67)	3.28	.001	9	.46
		9	BBS	MD 2.15 (−1.22 to 5.51)	1.25	.21	63	.67
		13	BI/MBI	SMD 0.30 (0.14–0.45)	3.73	<.001	0	1.00
	5–8 wk	4	FMA-UE	MD 6.02 (0.45–11.59)	2.12	.03	42	.46
		3	BBS	MD 0.60 (−1.88 to 3.08)	0.47	.64	0	.67
		2	BI/MBI	SMD 0.30 (−1.35 to 1.95)	0.36	.72	92	1.00
	≥9 wk	2	FMA-UE	MD 2.14 (−0.42 to 4.71)	1.64	.10	0	.46
		1	BBS	MD 2.80 (−2.99 to 8.59)	0.95	.34	–	.67
Weekly training volume	≤120 min/wk	3	FMA-UE	MD 2.47 (0.13–4.81)	2.07	.04	0	.41
		4	BBS	MD 1.31 (−1.25 to 3.86)	1.00	.32	17	.93
		4	BI/MBI	SMD 0.28 (−0.37 to 0.93)	0.85	.39	77	1.00
	121–299 min/wk	12	FMA-UE	MD 3.74 (1.75–5.74)	3.68	<.001	18	.41
		7	BBS	MD 1.71 (−1.59 to 5.00)	1.02	.31	31	.93
		9	BI/MBI	SMD 0.26 (0.09–0.44)	2.92	<.01	0	1.00
	≥300 min/wk	5	FMA-UE	MD 1.10 (−2.47 to 4.67)	0.60	.55	20	.41
		2	BBS	MD 3.89 (−9.93 to 17.71)	0.55	.58	86	.93
		1	BI/MBI	SMD 0.27 (−0.53 to 1.08)	0.67	.50	–	1.00
Total training dose	≤600 min	6	FMA-UE	MD 3.42 (1.22–5.62)	3.05	<.01	25	.47
		9	BBS	MD 1.56 (−0.65 to 3.77)	1.38	.17	23	.89
		10	BI/MBI	SMD 0.31 (0.14–0.48)	3.64	<.001	28	.15
	601–1200 min	8	FMA-UE	MD 3.61 (0.92–6.29)	2.63	<.01	12	.47
		3	BBS	MD 0.64 (−6.40 to 7.68)	0.18	.86	73	.89
		4	BI/MBI	SMD 0.03 (−0.31 to 0.37)	0.17	.87	44	.15
	>1200 min	6	FMA-UE	MD 1.63 (−0.81 to 4.07)	1.31	.19	0	.47
		1	BBS	MD 2.80 (−2.99 to 8.59)	0.95	.34	–	.89

Values are expressed as pooled effect sizes with 95% CIs. The within-subgroup P value refers to the test of the pooled effect within each dose subgroup. P for subgroup differences represents the interaction test comparing pooled effect estimates across dose subgroups for each outcome; P > .05 indicates no significant difference between dose subgroups.

– = not applicable, BBS = Berg Balance Scale, BI/MBI = Barthel Index/modified Barthel Index, CI = confidence interval, FMA-UE = Fugl–Meyer Assessment–Upper Extremity, I² = heterogeneity, MD = mean difference, SMD = standardized mean difference.

3.4.1. Session dose

In the subgroup analysis of session dose, upper limb motor function was not significantly improved when the session dose was >60 minutes (fixed-effect model, MD = 1.52, 95% CI: −2.29 to 5.33, P = .43). Conversely, ADL was significantly improved across all session doses, and the effect sizes were as follows: ≤30 minutes (fixed-effect model, SMD = 0.20, 95% CI: 0.04–0.37, P = .01); 31 to 60 minutes (fixed-effect model, SMD = 0.51, 95% CI: 0.15–0.87, P < .01). The balance function was not significantly improved under any session dose. Details are shown in Figure S1, Supplemental Digital Content 5.

3.4.2. Weekly training frequency

In the subgroup analysis of weekly training frequency, a weekly frequency ≥6 times did not significantly improve upper limb motor function (fixed-effect model, MD = 3.95, 95% CI: −1.82 to 9.71, P = .18). ADL was significantly improved only at a weekly frequency of 4 to 5 times (random-effects model, SMD = 0.28, 95% CI: 0.12–0.44, P < .001). The balance function was not significantly affected by any of the weekly training frequencies. Details are shown in Figure S2, Supplemental Digital Content 6.

3.4.3. Intervention duration

In the subgroup analysis of intervention duration, intervention duration ≥9 weeks failed to significantly improve upper limb motor function (fixed-effect model, MD = 2.14, 95% CI: −0.42 to 4.71, P = .10). For ADL, significant improvement was observed only in intervention duration ≤4 weeks (random-effects model, SMD = 0.30, 95% CI: 0.14–0.45, P < .001), but not in intervention duration of 5 to 8 weeks (random-effects model, SMD = 0.30, 95% CI: −1.35 to 1.95, P = .72). The balance function was not significantly affected by any intervention duration. Details are shown in Figure S3, Supplemental Digital Content 7.

3.4.4. Weekly training volume

In the subgroup analysis of weekly training volume, when the volume was ≥300 minutes, there was no significant improvement in upper limb motor function (fixed-effect model, MD = 1.10, 95% CI: −2.47 to 4.67, P = .55). For ADL, a significant improvement effect was only observed with weekly training volume of 121 to 299 minutes (random-effects model, SMD = 0.26, 95% CI: 0.09–0.44, P < .01). In terms of balance function, no significant improvement effect was found across all weekly training volumes. Details are shown in Figure S4, Supplemental Digital Content 8.

3.4.5. Total training dose

In the subgroup analysis of total training dose, when the total dose exceeded 1200 minutes, no statistically significant improvement in upper limb motor function was observed (fixed-effects model, MD = 1.63, 95% CI: −0.81 to 4.07, P = .19). For ADL, a statistically significant improvement was observed in the subgroup with a total training dose ≤600 minutes (fixed-effects model, SMD = 0.31, 95% CI: 0.14–0.48, P < .001). Balance function was not significantly affected by any total training dose (low, medium, or high doses). Details are shown in Figure S5, Supplemental Digital Content 9.

Exploratory subgroup analyses showed numerical differences in pooled effect estimates across different dose strata. For upper limb motor function, relatively larger pooled effect estimates were observed with a session dose ≤30 minutes, 4 to 5 sessions per week, an intervention duration of 5 to 8 weeks, a weekly training volume of 121 to 299 minutes, and a total training dose of 601 to 1200 minutes. For ADL, significant improvements were observed with a session dose of 31 to 60 minutes, 4 to 5 sessions per week, an intervention duration ≤4 weeks, a weekly training volume of 121 to 299 minutes, and a total training dose ≤600 minutes. No statistically significant improvement in balance function was observed in any dose subgroup. However, for FMA-UE, BBS, and BI/MBI, none of the subgroup difference tests for the 5 dose parameters reached statistical significance (all P > .05; Table 2). Therefore, although effect estimates differed across dose subgroups, these numerical differences should be interpreted cautiously because none of the subgroup interaction tests reached statistical significance.

3.5. Linear meta-regression of training dose

To further examine the linear dose–effect relationships of RAT, random-effects meta-regression analyses were performed separately for session dose, weekly training frequency, intervention duration, weekly training volume, and total training dose. None of the 5 dose parameters showed a statistically significant linear association with the treatment effects on FMA-UE, BBS, or ADL (all P > .05; Table S3, Supplemental Digital Content 3), and all 95% CIs for the regression coefficients included 0. Specifically, no significant linear association was observed for session dose, weekly training frequency, intervention duration, weekly training volume, or total training dose across any of the 3 outcomes. Overall, within the dose ranges represented by the included trials, no significant linear dose–effect relationship was identified between the examined training-dose parameters and functional outcomes.

3.6. Adjusted meta-regression analyses

The adjusted meta-regression analyses showed that, after controlling for time since stroke onset, none of the 5 training-dose parameters was significantly associated with treatment effects on FMA-UE, BBS, or ADL (all P > .05; Table S4, Supplemental Digital Content 4). In the BBS models, except for weekly training frequency (P = .220), time since stroke onset was significantly associated with treatment effects in the other 4 dose models (P = .008–.030), suggesting that it may be an important covariate influencing BBS treatment effects. However, after controlling for this covariate, none of the dose parameters showed a significant association with BBS treatment effects (P > .05), indicating that the dose–effect findings for BBS were relatively stable.

3.7. Restricted cubic spline analyses

Because the absence of a significant linear association does not exclude a potential nonlinear relationship, RCS analyses were further performed for session dose, intervention duration, weekly training volume, and total training dose. Three knots were placed at the 10th, 50th, and 90th percentiles of the observed distribution of each dose parameter. For FMA-UE, BBS, and ADL, none of the 4 dose parameters showed a statistically significant overall association with treatment effects, and none showed a statistically significant nonlinear association (all P > .05; Figs. S6–S9, Supplemental Digital Content 10). Specifically, both the overall association and nonlinearity tests were nonsignificant for session dose, intervention duration, weekly training volume, and total training dose across all 3 outcomes. Weekly training frequency was not modeled using RCS because its observed values were discrete and highly concentrated at 5 sessions/week. Overall, within the observed dose ranges, the RCS analyses provided no statistical evidence of nonlinear dose–effect relationships between the examined training-dose parameters and FMA-UE, BBS, or ADL.

3.8. Sensitivity analysis

For studies with low heterogeneity (I² < 50%), model-switching analysis was conducted, and the calculations were pooled using fixed-effect and random-effects models separately. Comparison of the 2 models revealed no significant changes in the effect sizes for the FMA-UE and BI/MBI outcome measures, indicating that the pooled results are generally reliable (Table 3).

Table 3.

Sensitivity analysis based on model switching.

Outcome measure	Fixed-effect model	Random-effects model	Consistency
FMA-UE (MD)	2.89 (95% CI: 1.49–4.28)	2.92 (95% CI: 1.39–4.46)	High
BI/MBI (SMD)	0.28 (95% CI: 0.13–0.43)	0.30 (95% CI: 0.10–0.50)	High

High consistency indicates that the direction, magnitude, and statistical significance of the pooled effect estimates were similar between the fixed-effect and random-effects models.

BI/MBI = Barthel Index/modified Barthel Index, CI = confidence interval, FMA-UE = Fugl–Meyer Assessment–Upper Extremity, MD = mean difference, SMD = standardized mean difference.

For studies with high heterogeneity (I² ≥ 50%), sensitivity analysis was conducted using the leave-one-out method. The sensitivity analysis results for the BBS outcome measure showed that removing any single study did not significantly affect the pooled effect size, indicating high robustness of the results (Fig. 7).

Figure 7.

Sensitivity analysis plot of the BBS outcome parameter. BBS = Berg Balance Scale, CI = confidence interval.

3.9. Publication bias

Funnel plots and Egger’s test were conducted separately for the 3 outcome parameters to assess publication bias. The funnel plots for FMA-UE and BI/MBI were generally symmetrical, although some individual study points fell outside the CIs (Fig. 8A and C). However, the Egger’s test results did not reach significance (FMA-UE: P = .675; BI/MBI: P = .275), indicating that no significant publication bias existed in these 2 outcome parameters. In contrast, the funnel plot for BBS was obviously asymmetric (Fig. 8B), and the Egger’s test result was significant (P = .018), suggesting a potential small-sample effect or publication bias in this outcome parameter.

Figure 8.

Funnel plot of publication bias. (A) FMA-UE; (B) BBS; (C) BI/MBI. BBS = Berg Balance Scale, BI/MBI = Barthel Index/modified Barthel Index, FMA-UE = Fugl–Meyer Assessment–Upper Extremity, WMD = weighted mean difference.

3.10. Adverse events

Among the 36 included studies, only 27 described adverse events. Among them, 19 studies explicitly reported no adverse events or no adverse events related to the intervention. A total of 8 studies (22.22%) reported actual adverse events, mainly including mild shoulder pain, skin abrasion, or occasional falls unrelated to training. Most adverse events were mild and transient. No serious adverse events directly related to RAT were observed.

3.11. Assessment of evidence quality

The levels of evidence for upper limb motor function and ADL were moderate, while that for balance function was low. The specific results are shown in Table 4.

Table 4.

GRADE quality assessment of evidence.

Outcome	Limitations of study	Inconsistency	Indirectness	Precision	Publication bias	Quality of evidence
Upper limb motor function	−1	0	0	0	0	Moderate
Balance function	−1	−1	0	0	−1	Low
ADL	−1	0	0	0	0	Moderate

Limitations of the studies: Multiple studies had a high risk of bias in randomization, allocation concealment, or blinding implementation, leading to downgrading. Inconsistency: Heterogeneity test showed I² ≥ 50%, downgraded. Indirectness: Weak association between outcome measures and assessment objectives, downgraded. Precision: Wide 95% CI for pooled effect size, downgraded. Publication bias: Asymmetric funnel plot and Egger’s test P value < .05, downgraded.

ADL = activity of daily living, CI = confidence interval, GRADE = Grading of Recommendations Assessment, Development and Evaluation.

4. Discussion

This meta-analysis included 36 studies and evaluated the effects of RAT on poststroke functional recovery based on 3 outcomes: upper limb motor function, balance function, and ADL, while also comparing treatment effects across different dose levels using 5 training-dose parameters. The overall analysis showed that RAT significantly improved upper limb motor function and ADL but had no significant effect on balance function. Exploratory subgroup analyses showed some numerical differences in pooled effect estimates across dose strata, but none of the between-subgroup differences were statistically significant. Further linear meta-regression identified no significant linear associations, and the RCS analyses likewise found no significant overall or nonlinear associations. Overall, RAT may improve some poststroke functional outcomes, but the current analyses did not identify a clear dose–effect relationship. These findings may provide a reference for future research on RAT dosing and the development of rehabilitation protocols.

The findings of the present study regarding the overall efficacy of RAT are consistent with those reported by Mehrholz et al^[54] and Baronchelli et al.^[55] RAT may improve upper limb motor function and ADL in patients with stroke to some extent, possibly because, compared with CR, it can provide continuous and stable high-repetition motor practice.^[56] Repetitive training is considered an effective approach for promoting poststroke recovery and facilitating neural network reorganization.^[38] In addition, RAT can maximize residual voluntary muscle activity while providing real-time visual performance biofeedback and proprioceptive feedback, thereby facilitating sensorimotor integration.^[48] These characteristics may contribute to the potential advantages of robotic therapy over conventional therapy. However, Hesse et al found that although RAT improved upper limb motor function in patients with stroke, it did not show a significant advantage over CR, suggesting that the effects of the 2 approaches may be comparable.^[26] In addition, some studies have suggested that the apparent advantages of RAT observed in certain trials may not be entirely attributable to robotic technology itself, but rather to longer training time and a greater training dose in the RAT group than in the CR group, which may affect treatment comparisons.^[47] Using a dose-matched meta-analysis, Boardsworth et al found a small but significant difference between RAT and CR in upper limb motor function, whereas no significant differences were observed in ADL or follow-up outcomes.^[57]

Regarding balance function, Loro et al found that the effects of RAT on balance were comparable with those of CR,^[58] which is consistent with the findings of the present study. Wang et al reported that RAT significantly improved balance function compared with CR; however, further secondary analysis indicated that this advantage was limited to patients in the acute or subacute stages of stroke, with no significant difference observed in the chronic stage.^[59] This finding differs from that of the present study, possibly because a relatively high proportion of the studies included in our review involved patients with chronic stroke. Similarly, Gandolfi et al reported in an RCT that RAT and CR produced comparable improvements in balance function among patients with chronic stroke.^[23] Amirbekova et al found in subgroup analyses that RAT appeared to be more effective in patients with acute or subacute stroke, whereas its effects were more limited in the chronic stage.^[9] Therefore, initiating intervention during the acute or subacute stage may be more favorable for balance recovery because the first 3 months after stroke are generally considered an important period for neuroplasticity-dependent recovery, during which recovery potential is relatively high.^[60] In addition, robotic devices differ in their assistance modes, control strategies, and feedback mechanisms. These device-related characteristics may influence postural control, sensory feedback, and active participation during training and may therefore affect improvements in balance function. Differences among robotic devices may thus represent 1 factor contributing to inconsistent findings across studies. At present, RAT may have beneficial effects on poststroke balance rehabilitation, but whether it provides additional benefits over CR requires further investigation.

Exploratory subgroup analyses showed numerical differences in pooled effect estimates across different dose strata, but none of the between-subgroup differences were statistically significant. Previous studies of RAT dosing have also produced inconsistent findings. Some studies have suggested that higher training intensity or dose may be associated with greater improvements in motor function,^[8] whereas others have found that simply increasing the training dose does not necessarily result in greater therapeutic effects, suggesting the possibility of a plateau effect.^[9] In the present study, none of the high-dose subgroups across the 5 dose dimensions reached statistical significance; however, because the between-subgroup differences were also nonsignificant, these high-dose subgroup findings should be interpreted cautiously. From a mechanistic perspective, neuroplasticity depends on intervention timing, duration, and individual patient characteristics, and its effects may be constrained by interindividual variability, the window of plasticity, and intrinsic recovery capacity. Once relevant neural networks have undergone effective reorganization, further increases in training volume may not consistently induce additional neuroplastic changes or functional improvements, potentially resulting in a plateau in treatment effects.^[61] In clinical practice, excessively high training doses may contribute to physical and cognitive fatigue in patients with stroke,^[62] which may subsequently reduce motivation and impair attention,^[63] ultimately affecting training quality and functional recovery. Moreover, rehabilitation outcomes depend not only on training dose but also on the specificity of the training content and the degree of active patient participation. Insufficient task specificity or excessive assistance resulting in passive participation may substantially reduce the actual training effect and may even adversely affect functional recovery.^[64,65]

More specifically, regarding upper limb motor function, Ko et al^[66] found that both session durations of ≤30 minutes and >30 minutes, as well as intervention durations of ≤2 weeks and >2 weeks, significantly improved FMA-UE scores. Huo et al^[67] also found that both total training times of ≤1000 min and >1000 min produced significant improvements, although the effect size was slightly larger in the lower-dose group (SMD: 0.67 vs 0.58). In contrast, the present study further subdivided higher-dose and longer-duration categories and found no significant improvement in the subgroups with session durations >60 minutes or intervention durations ≥9 weeks. This discrepancy may partly reflect the finer categorization of higher-dose ranges in the present study. Regarding ADL, Yang et al^[68] found that RAT significantly improved ADL at both 4 and 12 weeks, whereas in the present study, a significant improvement was observed only in the subgroup with an intervention duration ≤4 weeks, while the 5- to 8-week subgroup did not reach statistical significance. In addition to differences in the categorization of intervention duration, this discrepancy may also be influenced by differences in patient characteristics, robotic device types, and intervention protocols across the included studies. Regarding balance function, no significant improvement was observed in any dose subgroup across the 5 dose parameters in the present study, and none of the between-subgroup differences were statistically significant. This may partly reflect the fact that dose indicators such as training time, frequency, and total training volume do not fully capture actual training intensity. For example, even when training time is identical, differences in the level of robotic assistance and the degree of active patient participation may result in different actual training loads. Therefore, patients classified within the same nominal dose subgroup may still receive different levels of actual training stimulus, which may affect comparisons of treatment effects across dose subgroups.

In further analyses, none of the 5 training-dose parameters showed a significant linear association with FMA-UE, BBS, or ADL, and the RCS analyses likewise identified no significant overall or nonlinear associations. Previous studies examining dose–response relationships in RAT have generally been limited to upper limb function or a small number of dose-related variables. For example, Wu et al^[69] examined only total training time, number of training sessions, and average training time per session, and their meta-regression analyses similarly found no significant associations between these variables and improvements in upper limb motor function. The present study further examined the relationships between 5 dose parameters and 3 categories of functional outcomes and performed nonlinear analyses for dose parameters suitable for continuous modeling; nevertheless, no consistent dose-related associations were identified. This may partly be related to how training dose is quantified in the existing literature. Training duration, frequency, and total training volume mainly reflect the planned amount of training and cannot be considered equivalent to the actual training intensity or effective amount of practice completed by patients. Even when the reported training duration is the same, studies may differ in training intensity, task content, and the actual dose of training delivered.^[7,70] In addition, the distributions of some dose parameters were relatively concentrated, and only a limited number of studies represented the higher or lower ends of the dose range, which may have reduced the ability of linear meta-regression and RCS analyses to detect dose-related changes. Therefore, the current evidence does not provide sufficient support for a clear linear or nonlinear dose–effect relationship between RAT dose and rehabilitation outcomes, although a potential influence of RAT dose on poststroke rehabilitation cannot be excluded.

The dose-related findings of this study are also subject to several sources of clinical heterogeneity. Patients at different stages after stroke may differ in their recovery potential, and robotic devices vary in assistance modes, control strategies, and feedback mechanisms. In addition, the included studies comprised both RAT administered alone and RAT combined with CR, resulting in differences in actual treatment exposure. These sources of heterogeneity cannot be fully controlled in a study-level meta-analysis and may influence estimates of dose-related effects. Furthermore, some dose subgroups included only a limited number of studies, and none of the between-subgroup differences reached statistical significance; therefore, these subgroup findings should be regarded as exploratory. Future research should move beyond comparisons based on single dose indicators and incorporate more comprehensive assessments of actual training intensity, task content, and patient engagement to further determine whether a clear dose–effect relationship exists for RAT.

5. Limitations

This study has several limitations. First, some dose subgroups included only a small number of studies, and none of the between-subgroup differences reached statistical significance; therefore, these subgroup findings should be interpreted cautiously. Second, the 5 dose parameters were interrelated, making it difficult to distinguish the independent effects of each parameter. In addition, the values of some parameters were highly concentrated, which may have limited the ability of the meta-regression and RCS analyses to detect dose–effect relationships. Third, the included studies differed in stroke stage and robot type, and some patient characteristics could not be further adjusted for in the meta-regression analyses, which may have influenced the dose-related results. Fourth, existing studies provided limited information on actual training intensity and patient engagement; therefore, dose indicators such as training duration and frequency may not fully reflect the actual training dose. Finally, only English-language studies were included, which may have introduced language bias.

6. Conclusion

This study indicates that RAT may improve upper limb motor function and ADL after stroke, but has not yet shown a significant advantage in improving balance function. Current evidence does not demonstrate a clear linear or nonlinear relationship between training dose and rehabilitation outcomes. Future studies should adopt more standardized dose measurement and direct dose-comparison designs to further clarify the effects of different RAT training doses on poststroke rehabilitation outcomes.

Author contributions

Conceptualization: Song Hu, Daehee Kim.

Data curation: Song Hu, Fengjiao Wang, Jia Liu, Guoyu Wang.

Formal analysis: Song Hu.

Methodology: Song Hu, Jia Liu.

Supervision: Daehee Kim.

Validation: Fuxin Hong, Yulong He.

Visualization: Song Hu.

Writing – original draft: Song Hu.

Writing – review & editing: Song Hu, Jiaju Zhu, Yujie Chu, Daehee Kim.