Introduction

Cerebral palsy (CP) is a neurodevelopmental disorder caused by damage to the brain during early development and is characterized clinically by postural and motor dysfunction [1]. Motor dysfunction due to CP is often accompanied by sensory, perceptual, cognitive, communication, and behavioral deficits, as well as epilepsy and secondary musculoskeletal problems [2]. The different clinical features of CP can be divided into spastic, irregular, ataxic, and mixed CP [3]. The prevalence of CP is between 0.20 and 0.35% of surviving infants, and the number of people with CP is currently 17 million worldwide [4]. About 40% of these patients cannot walk independently, which seriously affects their activities of daily living [5]. Clinical rehabilitation interventions for CP patients mainly include neurodevelopmental therapy, physiotherapy, and hydrotherapy, which have achieved some success but are less effective in treating walking and balance functions [6].

Robot-assisted gait training (RAGT) is a new rehabilitation intervention that facilitates repetitive and efficient walking training with external mechanical assistance [7,8,9]. Lerner found that a robotic exoskeleton improved knee range of motion in patients with spastic CP [10]. Borggraefe trained 20 children with CP on a robot-assisted treadmill for 3 weeks, and they showed significant improvements in standing and walking function [11]. Digiacomo showed that RAGT combined with conventional treatment improved motor performance and endurance in children with CP [12].

However, the number of studies is small, outcome indicators and treatment effects vary, clinical effectiveness remains controversial, and there needs to be more high-quality evidence of evidence-based medicine to support this [13, 14]. Network meta-analysis is a statistical method that quantifies several interventions and prioritizes interventions’ effects based on different outcome indicators. In this study, we used the network meta-analysis method to search randomized controlled trials (RCTs) of RAGT for CP patients, extract relevant outcome indicators to evaluate their efficacy, and compare the differences in the efficacy of different robotic devices to provide evidence-based references for the clinical application of RAGT training.

Methods

This study was registered on the international system evaluation registration platform PROSPERO (CRD42022366471).

Search strategy

PubMed, Web of Science, Cochrane Library, Embase, CNKI, VIP, CBM, and Wanfang databases were searched by two researchers using medical mesh words and free words from the establishment of the database to December 26, 2022. The search formula is as follows: #1 “robot-assisted gait training” [MeSH] OR robot OR robotics OR RAGT, #2 “cerebral pals*” [MeSH] OR spastic quadriplegia OR spastic diplegia OR spastic hemiplegia OR CP, #3, #1, and #2. The condition was limited to a randomized controlled trial. Cross-check the results after the search, and discuss the decision with the third researcher in case of disagreement.

Inclusion criteria

The following are the inclusion criteria:

Study design—Chinese and English RCTs.

Participants—children aged <18 years with a precise diagnosis of CP, regardless of gender and race; the child is conscious, has a reasonable level of intelligence, and can cooperate with treatment and follow-up.

Intervention and control measures—the intervention group received RAGT combined with conventional rehabilitation, while the control group received conventional rehabilitation.

Outcome indicators—the primary outcome measures were Gross Motor Function Measure-88 (GMFM-88) D (standing and standing) and E (walking and running and jumping). The secondary outcome measures were the Berg Balance Scale (BBS), 6-minute walk test (6MWT), walking speed, and Modified Ashworth Scale (MAS).

Data—the experimental data are complete and provide sample sizes, means, and standard deviations within groups.

Exclusion criteria

Exclusion criteria are as follows: literature from non-randomized controlled trials; animal studies; conference abstracts, duplicate studies; incomplete outcome data; selective reporting studies; unavailable full text; literature not in Chinese or English.

Study selection and data extraction

Two researchers carried out the search, screening, and collation of the literature and the extraction of data and information independently, and the results were cross-checked after completion. In case of disagreement, a third researcher was assigned to decide. Data extraction included the first author, year of publication, basic information about the study population (sample size, age), basic information about the interventions (interventions, duration of treatment, follow-up, type of robot), and outcome indicators.

Quality assessment

Two researchers assessed the risk of bias in the included literature according to the Cochrane Risk Assessment Scale and the Physiotherapy Evidence Database (PEDro) scale. The third researcher discussed the decision in any disagreement between the two researchers.

Statistical analysis

RevMan 5.4 was used for meta-analysis. The data types of the outcome indicators for this meta-analysis were all measures. The standard mean difference (SMD) and its 95% confidence interval (CI) were used for statistical analysis of effect values. A fixed effect model was used to analyze if I2 ≤ 50% and P ≥ 0.1, indicating a small heterogeneity among the included studies. A random effect model was used for analysis if I2 > 50%, P < 0.1, indicating more significant heterogeneity between studies. Subgroup analysis or sensitivity analysis was used to identify sources of heterogeneity. The significance level α=0.05.

R 4.2.0 was used to make the network relationship map of the efficacy comparison of different robotic devices. ADDIS 1.16.8 was used for network meta-analysis. Since there is no closed loop in this study’s evidence relationship network graph, a consistency model was fitted to perform a probabilistic analysis of the outcome indicators. Four Markov chains were set, the initial value of the model was 2.5, the iteration step was fine-tuned by 10, and the number of iterations was 50,000, of which the first 20,000 were annealed to eliminate the effect of the initial value. The last 30,000 were used for sampling. The potential scale reduction factor (PSRF) was used to evaluate the convergence efficiency. The PSRF tends to be 1, indicating good convergence efficiency and high reliability of the network meta-analysis results. The rank-ranking probability diagram presented the probability of each intervention becoming the best intervention, and rank 1 was the best probability. The ranking of interventions’ superiority was based on the magnitude of the rank 1 value. The funnel plot was drawn using RevMan 5.4, and if the funnel plot was symmetrical, it suggested that the possibility of publication bias was low.

Results

Study selection and characteristics

According to the literature search strategy, 428 studies were obtained from the initial search. After excluding 192 duplicate publications, the remaining 236; after reading the titles and abstracts, 204 were excluded, and the remaining 32. After carefully reading the complete text, 14 studies were finally included, including 8 in Chinese [15,16,17,18,19,20,21,22] and 6 in English [23,24,25,26,27,28], with a total of 654 patients. The process of literature screening is shown in Fig. 1. The essential characteristics of the included studies are shown in Table 1.

Fig. 1
figure 1

Screening process of literature selection

Full size image
Table 1 Characteristics of included studies
Full size table

Quality and risk of bias assessment

All studies mentioned “random” or “randomized grouping,” but two studies [21, 23] did not specify the specific randomized scheme. Two studies [25, 27] used opaque envelopes to complete allocation hiding, while the remaining studies did not mention allocation sequence hiding. One study [24] blinded the treatment, and four studies [23,24,25, 28] were blinded to outcome assessors. The data results included in the literature were complete and not reported selectively. One study [26] had another bias. Nine included studies [16, 18, 19, 23,24,25,26,27,28] were of high quality, and five [15, 17, 20,21,22] were of moderate quality, with a mean score of 7.14. The results of the quality assessment of the included studies are shown in Fig. 2 and Table 2.

Fig. 2
figure 2

Risk assessment of bias

Full size image
Table 2 Physiotherapy evidence database scores of the included studies
Full size table

Meta-analysis

GMFM-88 D score

Nine studies [16, 18,19,20,21,22, 24, 26, 27] reported GMFM-88 D scores. Heterogeneity tests suggested significant heterogeneity among studies (I2 = 57%, P = 0.02), and a random effect model was used to combine effect sizes. Overall, the GMFM-88 D score was higher in the experimental group than in the control group (SMD = 0.84, 95% CI 0.54 to 1.15, P < 0.05) (Fig. 3). Subgroup analysis showed that GMFM-88 D scores were significantly higher in both the LokoHelp and Lokomat groups than in the control group (SMD = 1.14, 95% CI 0.89 to 1.40, P < 0.05 and SMD = 0.71, 95% CI 0.20 to 1.23, P < 0.05). However, the difference between the 3DCaLT and control groups was insignificant (SMD = −0.11, 95% CI −0.97 to 0.75, P > 0.05).

Fig. 3
figure 3

Effects of RAGT on GMFM-88 D scores of patients with CP

Full size image

GMFM-88 E score

Nine studies [16, 18,19,20,21,22, 24, 26, 27] reported GMFM-88 E scores. Heterogeneity tests suggested significant heterogeneity among studies (I2 = 68%, P = 0.002), and a random effect model was used to combine effect sizes. Overall, GMFM-88 E scores were significantly higher in the experimental group than in the control group (SMD = 0.78, 95% CI 0.43 to 1.14, P < 0.05) (Fig. 4). Subgroup analysis showed that GMFM-88 E scores were higher in both the LokoHelp and Lokomat groups than in the control group (SMD = 0.88, 95% CI 0.31 to 1.46, P < 0.05 and SMD = 0.80, 95% CI 0.28 to 1.32, P < 0.05). However, there was no significant difference between the 3DCaLT and control groups (SMD = 0.13, 95% CI −0.73 to 0.99, P > 0.05).

Fig. 4
figure 4

Effects of RAGT on GMFM-88 E scores of patients with CP

Full size image

BBS

Five studies [15,16,17,18, 22] reported BBS scores. Heterogeneity tests suggested significant heterogeneity among studies (I2 = 71%, P = 0.009), and a random effect model was used to combine effect sizes. Overall, BBS scores were higher in the experimental group than in the control group (SMD = 0.91, 95% CI 0.50 to 1.32, P < 0.05) (Fig. 5). Subgroup analysis showed that BBS scores were higher in both the LokoHelp and Lokomat groups than in the control group (SMD = 1.00, 95% CI 0.51 to 1.50, P < 0.05 and SMD = 0.57, 95% CI 0.09 to 1.04, P < 0.05).

Fig. 5
figure 5

Effects of RAGT on BBS scores of patients with CP

Full size image

6MWT

Three studies [21, 25, 27] reported the results of the 6MWT evaluation. Heterogeneity tests suggested no significant heterogeneity among studies (I2 = 0%, P = 0.75), and a fixed effect model was used to combine effect sizes. Overall, the results of the 6MWT evaluation were higher in the experimental group than in the control group (SMD = 0.67, 95% CI 0.18 to 1.15, P < 0.05) (Fig. 6). Subgroup analysis showed that 6MWT evaluation results were higher in the 3DCaLT group than in the control group (SMD = 0.91, 95% CI 0.00 to 1.82, P = 0.05). 6MWT evaluation results in both the Lokomat and Gait Trainer 1 group were not statistically significant compared to the control group (SMD = 0.66, 95% CI −0.07 to 1.40, P > 0.05 and SMD = 0.41, 95% CI −0.53 to 1.35, P > 0.05).

Fig. 6
figure 6

Effects of RAGT on 6MWT of patients with CP

Full size image

Walking speed

Four studies [17, 23, 26, 27] reported walking speed (m/s) in patients with CP after treatment. Heterogeneity tests suggested no significant heterogeneity among studies (I2 = 0%, P = 0.49), and a fixed effect model was used to combine effect sizes. Overall, the walking speed of patients in the experimental group was higher than that of the control group. The difference between the experimental and control groups was not statistically significant (SMD = 0.31, 95% CI −0.08 to 0.71, P > 0.05) (Fig. 7).

Fig. 7
figure 7

Effects of RAGT on the walking speed of patients with CP

Full size image

MAS

Five studies [15, 17, 18, 27, 28] reported MAS scores. Heterogeneity tests suggested significant heterogeneity among studies (I2 = 93%, P < 0.00001), and a random effect model was used to combine effect sizes. Overall, the MAS scores were lower in the experimental group than in the control group. The difference between the experimental and control groups was not statistically significant (SMD = −0.67, 95% CI −1.75 to 0.41, P > 0.05) (Fig. 8).

Fig. 8
figure 8

Effects of RAGT on MAS scores of patients with CP

Full size image

Network meta-analysis

Evidence network

The GMFM-88 E score, more closely related to lower limb motor function, was used to compare the efficacy of different robotic. Of the studies reporting GMFM-88 E scores, four studies [16, 18,19,20] used the LokoHelp rehabilitation robot (RECK Company), four studies [21, 22, 24, 26] used the Lokomat children's version rehabilitation robot (HOCOMA Company), and one study [27] used 3DCaLT robot device. The network relationship is shown in Fig. 9 (a, LokoHelp; b, Lokomat; c, 3DCaLT; d, control).

Fig. 9
figure 9

The network relationship of efficacy Comparison of different robot

Full size image

Consistency test

There was no closure of the loop between interventions in this study. Therefore, consistency testing was not required.

Convergence diagnosis

A convergence diagnosis of the included studies revealed bandwidth values close to 0, suggesting good convergence (Fig. 10).

Fig. 10
figure 10

Track density map of the GMFM-88 E score

Full size image

Probability ranking

The network meta-analysis revealed that the best ranking of probabilities for the effect of 3 different robots on GMFM-88 D scores was LokoHelp (P = 0.66) > Lokomat (P = 0.28) > 3DCaLT (P = 0.06) and the best ranking of probabilities for the GMFM-88 E score was LokoHelp (P = 0.63) > 3DCaLT (P = 0.21) > Lokomat (P = 0.16) (Fig. 11, Fig. 12, Table 3).

Fig. 11
figure 11

Rank of GMFM-88 D scores

Full size image
Fig. 12
figure 12

Rank of GMFM-88 E scores

Full size image
Table 3 Best probability ranking
Full size table

Adverse reaction

One study [17] reported that RAGT training caused a small number of patients to experience mild discomfort from rubbing the skin on the inner root of the thigh from wearing the robot device, which was relieved by rest and did not interfere with the next day’s treatment. No adverse effects were reported in other studies.

Subgroup analysis of primary outcomes

GMFM-88 D and E scores were analyzed in subgroups according to the intervention time and patient age. The analysis results generally follow the above analysis (Table 4).

Table 4 Subgroup analysis of RAGT on patients with CP
Full size table

Sensitivity analysis

The meta-analysis results were analyzed for sensitivity using a one-by-one exclusion method, removing one study at a time. The results showed no significant change from the above results, indicating that the meta-analysis results were relatively stable.

Publication bias

No significant asymmetry was seen in the funnel plot of the study with GMFM-88 D and E scores as outcome indicators. It is suggested that the results of data analysis are less influenced by publication bias (Fig. 13).

Fig. 13
figure 13

Funnel plot

Full size image

Discussion

Lower limb motor dysfunction is one of the most common symptoms in patients with CP [29, 30], leading to an inability to walk or abnormal gait patterns, such as horseshoe foot or crouching gait [31]. It can cause secondary injuries and have a severe impact on the quality of life of patients. Therefore, the balance of lower limbs and walking functions are essential in rehabilitating patients with CP [32, 33]. RAGT, as an emerging rehabilitation therapy, brings new hope for the functional rehabilitation of CP patients [34]. This study analyzed the literature on RAGT for lower limb motor function in patients with CP using a network meta-analysis. We compared the clinical efficacy of different robotic devices.

The meta-analysis showed that RAGT training improved GMFM-88 D and E area scores, BBS scores, and 6MWT significantly better than conventional rehabilitation in CP patients. However, the effect of improved gait speed and lower limb muscle spasticity is unclear. The network meta-analysis results showed that RAGT improved lower limb motor function in CP patients, with LokoHelp most likely to improve lower limb motor function to a greater extent in CP patients.

Central nervous system plasticity theory and motor learning theory, closely related to the reconstruction of brain cell function, suggest that repetitive movements can induce plastic changes in the brain. Neural plasticity changes are primarily realized by rebuilding neural networks [35,36,37,38]. The therapeutic mechanism of RAGT on lower limb motor function in patients with CP may be related to the improvement of nervous system plasticity [39]. Shin showed that repetitive RAGT training promotes functional network reorganization in the brain’s motor cortex and stimulates neuroplasticity [40]. Youssofzadeh found that RAGT training significantly improved functional network connectivity around the sensory-motor areas of the brain and enhanced functional connectivity between the prefrontal and parietal lobes after EEG analysis in healthy subjects [41].

The following limitations exist in this study: the total number of subjects included in the literature was small; some of the literature was not allocation sequence concealed and blinded, which may lead to a potential risk of bias; there were age differences in the study population and varying severity of the condition, which may have had an impact on the rehabilitation outcome; the frequency and periodicity of interventions varied in the literature, which may have biased the findings; some of the literature had short treatment periods.

Conclusion

In conclusion, RAGT training improved the motor function of the lower limbs of CP patients, with LokoHelp probably having the best treatment effect. This positively impacted the patient’s quality of life, balance function, and motor control. However, the evidence for the efficacy of RAGT in improving gait speed and muscle spasticity is insufficient and needs further exploration. In addition, the small sample size of the literature included in this study and the variable quality of the articles may have impacted the analysis results. Therefore, adequate sample size and high-quality and long-term follow-up studies are still needed to explore the clinical efficacy of RAGT in depth.