Robot-assisted gait training on functional ambulation in patients with stroke: a retrospective real-world analysis across five robotic devices Authors: Dougho Park, Dahyeon Koo, Gwang-Min Park, Won Hyuk Chang, Yong Wook Kim, Seo Yeon Yoon, Deog Young Kim, Dae Hyun Kim Published: 2026 ABSTRACT AbstractBackgroundRobot-assisted gait training (RAGT) is increasingly used in stroke rehabilitation. However, its comparative effectiveness across device types remains unclear. This study explored the effectiveness of RAGT compared to conventional treatment using real-world clinical data.MethodsThis retrospective cohort study included 2,947 patients with stroke who received inpatient rehabilitation at a single center between 2015 and 2022. Patients were divided into RAGT and conventional treatment groups. Functional improvement was defined as a gain of ≥ 1 point in the Functional Ambulation Category. Multivariable logistic regression and classification and regression tree (CART) models were applied to identify predictors of favorable outcomes (effectors) and to explore clinical decision-making for device selection.ResultsThe RAGT group showed poorer baseline function but had a similar proportion of effectors (58.0%) compared to the control group (62.5%). Overall, RAGT was not associated with effector outcome versus control. Among RAGT devices, Morning Walk (adjusted odds ratio [aOR] = 3.25; 95% confidence interval [CI]: 1.79–5.87) and Exowalk (aOR = 4.42; 95% CI: 1.84–10.64) were significantly associated with higher odds of effector status compared to control, whereas Erigo was associated with lower odds (aOR = 0.58; 95% CI: 0.35–0.96). CART models identified clinical features, such as the Berg Balance Scale, the Mini-Mental Status Examination, and the Fugl-Meyer Assessment, as key predictors of effector status and revealed patterns in robotic device allocation.ConclusionWhile RAGT was not superior to conventional treatment, certain devices were associated with improved gait function. These findings highlight the importance of tailoring robotic interventions.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12984-026-02010-w.Keywords: Ambulatory care, Gait disorders, Recovery of function, Robotics, Stroke rehabilitation Background Stroke is a leading cause of adult disability worldwide, and gait impairment is one of its most debilitating consequences [1]. Regaining independent walking is a central goal in post-stroke rehabilitation, but conventional gait training, such as therapist-assisted overground or treadmill walking, has inherent limitations. Although conventional treatment can improve mobility, it is labor-intensive, restricted by therapist fatigue, and often limited in both intensity and duration [2]. These challenges have led to the development of robot-assisted gait training (RAGT) devices, which use electromechanical support to enable high-repetition, task-specific walking practice [3]. RAGT can deliver high-dose, intensive gait training while facilitating task-specific, repetitive lower limb movement that approximates the kinematic parameters of normal walking, although the degree to which each device reproduces a physiological gait pattern varies by device-specific mechanical design [4]. Robotic systems allow early gait rehabilitation even for patients with severe deficits who cannot yet support themselves, providing safe, upright mobilization with reduced fall risk and decreased physical burden on therapists [2]. By automating and augmenting gait practice, RAGT offers a means to intensify rehabilitation beyond what manual therapy alone can achieve [5]. A 2020 Cochrane review involving 62 trials with 2,440 participants showed that RAGT combined with conventional physiotherapy significantly increased the likelihood of achieving independent walking, particularly for patients in the first three months post-stroke and those unable to walk independently [2]. A recent meta-analysis of 24 randomized controlled trials confirmed that RAGT significantly improved lower extremity motor function, balance, walking ability, and endurance compared to conventional gait training [1]. However, several critical limitations remain in the existing evidence base. First, most randomized controlled trials have been conducted under highly controlled research conditions with strict inclusion and exclusion criteria, which may limit their generalizability to real-world clinical practice. Second, the comparative effectiveness of different robotic devices remains unclear due to limited head-to-head comparisons and lack of standardized guidance for device selection [6]. Third, few studies have explored how patient-specific characteristics should inform robotic device selection, despite theoretical frameworks suggesting that device allocation should reflect patients’ functional status [7, 8]. Finally, previous research has primarily focused on device efficacy rather than effectiveness, offering limited insight into how RAGT performs in routine clinical settings with diverse patient populations. To address these real-world gaps, our study examined five robotic gait-training devices available at our institution, Severance Hospital, Seoul, South Korea. These devices were selected as institution-specific examples spanning different positions within the established framework for robotic gait training, rather than as a proposed universal taxonomy. RAGT devices are broadly classified into two main types: exoskeleton devices, which apply forces directly to lower-limb segments, and end-effector devices, which guide movement through contact with the feet [2, 9]. A further distinction is made between treadmill-based and overground systems based on ambulation mode [10]. The Lokomat (Hocoma AG, Volketswil, Switzerland) is a treadmill-based exoskeleton that drives hip and knee movements through a rigid orthosis with harness-based body weight support (BWS) [11]. The Exowalk (HMH Co. Ltd, Incheon, South Korea) is an overground wearable exoskeleton that provides powered joint assistance during free walking without a fixed frame or overhead harness [12]. The Morning Walk (Curexo Inc., Seoul, South Korea) is a treadmill-based end-effector device that guides foot trajectory through motorized footplates combined with a saddle-seat support system [13]. The Erigo (Hocoma AG, Volketswil, Switzerland) is a tilt-table–based end-effector device that combines progressive verticalization with dynamic leg movement, primarily indicated for non-ambulatory patients in the early rehabilitation phase [14]. The Andago (Hocoma AG, Volketswil, Switzerland) is an overground BWS device that provides harness-based partial body weight support and fall prevention during self-initiated walking without applying active joint-driven assistance [15]. This mechanical and functional diversity suggests that device effectiveness may depend not only on stroke severity but also on the complex interplay among a patient’s motor control capacity, postural stability, and cognitive function. However, this hypothesis requires validation in real-world settings through a comprehensive analysis of patient-device matching and clinical outcomes. This study addresses these critical knowledge gaps by analyzing real-world data from a tertiary-level university-affiliated robotic rehabilitation center that functions as a regional referral center, receiving post-acute stroke patients from primary and secondary care institutions, over seven years. Unlike previous research conducted under controlled trial conditions, our study investigates how clinicians assign patients to one of five robotic devices based on comprehensive functional assessments. More importantly, we identify which patients can tolerate and benefit from specific RAGT devices in everyday clinical practice. Our aim was not to propose a universal five-category taxonomy of RAGT, but to compare the real-world clinical effectiveness, patient suitability, and practical application differences among the five robotic gait-training devices used in our institutional program. Through retrospective analysis of 2,947 stroke patients, we investigated three primary research questions: (1) what are the real-world clinical selection criteria for different RAGT devices based on functional status, (2) does RAGT demonstrate overall superiority compared to conventional gait training in a broad, clinically unselected cohort of patients with stroke receiving routine inpatient rehabilitation, and (3) which patient-specific factors predict functional ambulation improvement and how do these factors influence the comparative effectiveness of different RAGT devices. This study uniquely combines large-scale clinical data with decision-tree modeling to establish evidence-based guidelines for patient-robot matching, thereby bridging the gap between efficacy trials and clinical implementation. Methods Study design and patient selection This retrospective cohort study analyzed data from patients with stroke who received inpatient rehabilitation between January 2015 and April 2022 at a tertiary-level university-affiliated rehabilitation center. We included patients who completed a 6-week inpatient rehabilitation program according to the routine institutional protocol and had a stroke confirmed by computed tomography or magnetic resonance imaging. Patients were divided into two groups: the robot-assisted group, who received RAGT in addition to conventional physical therapy, and the conventional treatment group, who received only conventional physical therapy during their inpatient stay. The decision to assign a patient to the RAGT or control group was not randomized; rather, it was based on clinical judgment by the treating rehabilitation team, informed by functional status at admission, availability of robotic devices during the study period, and patient-specific factors such as tolerance and safety. A total of 2,956 patients were initially identified. After excluding patients with duplicate records (n = 7) and those with incomplete medical documentation regarding demographic characteristics or primary diagnosis (n = 2), 2,947 patients remained for analysis, including 1,618 in the conventional treatment group (control group) and 1,329 in the RAGT group (Fig. 1). Flowchart of patient inclusion. RAGT, robot-assisted gait training The institutional review board of Severance Hospital approved the procedures and protocols of this study on June 17, 2022 (approval no. 4-2022-0546). Owing to the retrospective design, the requirement for informed consent was waived. Robot-assisted gait training The RAGT program at our institution began in April 2011 with the introduction of the Lokomat V5. In September 2018, the rehabilitation department expanded its robotic system by acquiring the upgraded Lokomat V6, as well as Erigo and Andago. The system was further enhanced in September 2019 with the addition of Erigo Pro, Morning Walk, and Exowalk. The characteristics of the RAGT devices used in this study are presented in Table S1. Each RAGT session was conducted on weekdays, five days per week, and designed to provide 30 min of active robotic training, excluding the time required for donning and doffing. For patients unable to tolerate the full 30-minute session, training duration was adjusted based on individual tolerance. The selection of the initial robotic device was made by a team of three neurorehabilitation specialists using a pragmatic clinical matching approach rather than a fixed numeric algorithm. In general, the clinicians aimed to prescribe the highest task-demand, least-constraining robotic system that was judged to be safe and tolerable for the patient, while considering baseline ambulation level, trunk control and balance, lower-limb motor impairment, ability to maintain upright posture, cognitive ability to follow instructions, and overall safety during gait training. This approach was informed by the conceptual framework proposed by Morone et al. [7], which emphasizes matching patient ability and gait deficit to the degree of machine constraint and achievable training intensity. Because tolerability could be fully verified only during actual training, device suitability was reassessed during the first few sessions, and the initially selected device was maintained or changed according to observed tolerance and performance. If a patient showed improvements in gait function, lower limb strength, or balance, suggesting a better-suited device, a transition protocol was initiated: the patient used the initial device for 10 sessions and then switched to the more suitable device for the remaining 10 sessions. The standard RAGT protocol consisted of 20 sessions over four weeks as part of a six-week comprehensive inpatient rehabilitation program. Within each device, the available settings were individualized by the rehabilitation team to maintain safe participation at the highest tolerable challenge level. Depending on the platform, these settings could include body-weight support/unloading, gait speed, robotic assistance/guidance, step-related parameters, and tilt/verticalization level. These device-specific parameter values were not consistently stored in the retrospective database and were therefore not included in the present analyses. Variable and outcome definitions Baseline patient characteristics included age, sex, and body mass index, which are established confounders of functional recovery after stroke [16]. Stroke type was categorized as ischemic stroke, hemorrhagic stroke, or stroke with other etiologies, given evidence that these subtypes differ in initial severity, recovery trajectory, and rehabilitation outcomes [17]. Comorbidities included hypertension, diabetes mellitus, dyslipidemia, and atrial fibrillation, as recorded in the patient’s medical history, as these conditions may independently affect rehabilitation potential [16, 18]. Patients were assessed before the first session and after the last session of each intervention. Total sessions completed were also recorded. Outcome measures included the Functional Ambulation Category (FAC), Fugl-Meyer Assessment for Lower Limb (FMLL), Modified Barthel Index (MBI), Berg Balance Scale (BBS), Mini-Mental State Examination (MMSE), and Medical Research Council (MRC) scale for bilateral hip flexors, knee extensors, and ankle dorsiflexors; these functional scales were selected to capture the multidimensional nature of stroke impairment–motor function, balance, cognition, activities of daily living, and ambulation–each of which has established validity in post-stroke rehabilitation research [19, 20]. In the RAGT group, we additionally recorded the number of days from stroke diagnosis to the first RAGT session. The primary outcome was functional ambulation improvement, defined as a ≥ 1-point increase in FAC score from baseline to follow-up. This threshold was selected based on the ordinal structure of the FAC, in which each category transition represents a clinically significant change in the level of physical assistance required for ambulation [21]. Furthermore, a ≥ 1-point FAC change has been used as an anchor for defining clinically meaningful improvement in prior stroke rehabilitation research [22, 23]. Patients meeting this criterion were classified as effectors; those who did not were classified as non-effectors. Patients with an initial FAC score of 5 were excluded from this analysis to avoid ceiling effects. Statistical analysis Before analysis, data integrity was verified through the following procedures: duplicate records were identified and removed (n = 7), and records with incomplete demographic or diagnostic documentation were excluded (n = 2), as described in the patient selection section. Range checks were performed on all continuous variables to identify biologically implausible values. No additional records were excluded on this basis, as all values were confirmed to represent valid clinical observations upon source verification. Continuous variables are presented as means ± standard deviations, and categorical variables as frequencies and proportions. Normality of all continuous variables was assessed using the Shapiro-Wilk test (all p-values < 0.001). Nevertheless, given the large sample size (n = 2,947) and consistency of results across both parametric and non-parametric approaches (verified by parallel Mann-Whitney U tests), independent t-tests were retained for two-group comparisons of continuous variables. For comparisons across three or more groups, one-way ANOVA was used for continuous variables; results were verified using the Kruskal-Wallis test, which yielded consistent conclusions for all variables. Categorical variables were compared using the Pearson chi-square test, with chi-square trend tests applied to ordinal variables where appropriate. Within-group pre-to-post changes in FAC were assessed using the Wilcoxon signed-rank test, given the ordinal structure of the scale. Independence of observations at the patient level was ensured by the single-episode-per-patient study design. Multivariable logistic regression models were constructed to identify predictors of favorable outcome (effector status). Patients with an initial FAC score of 5 were excluded to reduce ceiling effects. In the overall cohort model, covariates included age, sex, body mass index, stroke type, comorbidities (hypertension, diabetes mellitus, dyslipidemia, atrial fibrillation), baseline functional scores (FMLL, MBI, MMSE, BBS, MRC, and FAC), total number of sessions completed, and treatment group (RAGT vs. Control groups). In the RAGT-only model, the same covariates were included, along with the number of days from diagnosis to the first RAGT session. Multicollinearity among covariates was assessed using the adjusted generalized variance inflation factor (GVIF). We defined adjusted GVIF < 5 as acceptable. To evaluate the influence of extreme values ​​on study conclusions, a pre-specified sensitivity analysis was conducted after excluding patients with session counts exceeding 200 (n = 17), body mass index (BMI) below 15 kg/m² (n = 16), or BMI above 40 kg/m² (n = 5), as these were all considered to be variables with a significant influence on walking results; a total of 28 patients after accounting for overlapping exclusion criteria. All primary analyses were repeated in this restricted sample, and results were compared with those of the primary analysis to assess robustness. To examine whether the sequential introduction of new devices influenced clinical outcomes within the Lokomat group, a post-hoc subgroup analysis was conducted by stratifying Lokomat patients according to enrollment period relative to the September 2019 device portfolio expansion, when all five RAGT devices became available at our institution. Patients enrolled before September 2019 were designated as the Lokomat-early group, and those enrolled thereafter as the Lokomat-late group. Baseline characteristics and effector rates were compared between these two subgroups. Additionally, each subgroup was independently compared with the control group using the same multivariable logistic regression framework as the primary device-specific analyses. We also employed classification and regression tree (CART) models using the “rpart” package to predict effector status and initial robotic device assignment [24]. Binary CART models used FMLL (affected side), FAC, BBS, and MMSE as predictors. A separate multiclass CART model was applied to identify robot device assignment among effectors. Trees were constructed using Gini impurity, with tuning of complexity parameters (cp.), maximum depth, and minimum node size to enhance interpretability and reduce overfitting. Overfitting was addressed through three complementary measures: (1) pre-specification of tree complexity parameters (binary CART: cp. = 0.006, maximum depth = 6, minimum node size = 20; multiclass CART: cp. = 0.01, maximum depth = 5, minimum node size = 10); (2) selection of the final tree based on the minimum cross-validated error derived from the built-in 10-fold cross-validation implemented in the rpart package; and (3) comparison of training error and cross-validated error to assess the degree of overfitting. The performance of the binary CART model was evaluated using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and the area under the receiver operating characteristic curve (AUC). These performance metrics were calculated on the derivation dataset and should be interpreted as apparent performance (training dataset). The multiclass CART model was constructed as an exploratory analysis to identify patterns in clinical device assignment among effectors, rather than as a predictive classification tool; accordingly, its results should be interpreted as hypothesis-generating. Final trees were visualized using “rpart.plot” with class probability bars [25]. A p-value of < 0.05 was considered statistically significant. Missing data were not imputed, and all regression analyses were conducted on complete cases, defined as patients with no missing values on any covariate included in the respective model. Of the 2,947 patients enrolled, 381 with an initial FAC score of 5 were excluded to avoid ceiling effects, and 585 were excluded due to missing follow-up FAC data, yielding a primary analysis set of 1,981 patients. An additional 475 patients were excluded from the logistic regression models due to missing values in one or more covariates, primarily FMLL (11.2%), MMSE (7.9%), and BBS (6.6%), resulting in a final analytic sample of 1,506 patients. Statistical computations were performed using R software (version 4.4.1). The analysis code is available as online supplementary material. Results Baseline characteristics Compared with the control group, the RAGT group was significantly older (p = 0.007) and had higher proportions of patients with comorbidity (all p < 0.001). The distribution of stroke types also differed, with the RAGT group having a relatively higher proportion of hemorrhagic stroke (p = 0.002). Regarding initial assessments, the RAGT group had significantly lower baseline FAC scores (p < 0.001) and worse performance in FMLL (p < 0.001), MBI (p < 0.001), MMSE (p = 0.004), and BBS scores (p = 0.002). Among these assessments, FMLL on the lesion side had the largest effect size (Cohen’s d = 0.424). The number of sessions was significantly higher in the control group (37.4 ± 48.8) than in the RAGT group (21.2 ± 44.0, p < 0.001) (Table 1). Table 1Baseline characteristics of the overall study populationVariablesGroupsp-valueEffect sizeaControl (n = 1618)RAGT (n = 1329)Age, years60.7 ± 15.562.3 ± 15.10.0070.101Male, n (%)907 (56.1)769 (57.6)0.4100.016Body mass index, kg/m222.8 ± 3.622.9 ± 3.90.3850.033Primary diagnosis, n (%)0.0020.065Ischemic stroke809 (50.1)603 (45.4)Hemorrhagic stroke746 (46.1)692 (52.1)Others62 (3.8)34 (2.6)Hypertension, n (%)1046 (64.6)1047 (78.8)< 0.0010.155Diabetes, n (%)393 (24.3)445 (33.5)< 0.0010.101Dyslipidemia, n (%)123 (7.6)319 (24.0)< 0.0010.229Atrial fibrillation, n (%)211 (13.0)247 (18.6)< 0.0010.076FAC, n (%)< 0.0010.1660642 (47.6)768 (58.9)1200 (14.8)224 (17.2)2123 (9.1)96 (7.4)3183 (13.6)123 (9.4)4133 (9.9)74 (5.7)568 (5.0)18 (1.4)Initial FMLLb20.9 ± 10.016.7 ± 9.9< 0.0010.424Initial MBI30.6 ± 30.025.0 ± 25.6< 0.0010.197Initial MMSE13.4 ± 7.612.5 ± 8.30.0040.114Initial BBS12.7 ± 17.410.8 ± 14.30.0020.118Initial MRCc13.7 ± 7.316.5 ± 6.3< 0.0010.414Sessions, n37.4 ± 48.821.2 ± 44.0< 0.0010.347Effectorsd, n (%)487 (62.5)697 (58.0)0.0500.045aCalculated by Cohen’s d for continuous variables and Cramér’s V for categorical variablesbAt the lesion sidecfor bilateral hip flexor, knee extensor, dorsiflexor musclesdn = 1281:1285 (excluding cases with an initial FAC = 5)BBS, Berg balance scale; FAC, functional ambulation category; FMLL, Fugl-Meyer assessment for lower limb; MBI, modified Barthel index; MMSE, mini-mental state examination; MRC, Medical Research Council scale; RAGT, robot-assisted gait training Baseline characteristics of the overall study population aCalculated by Cohen’s d for continuous variables and Cramér’s V for categorical variables bAt the lesion side cfor bilateral hip flexor, knee extensor, dorsiflexor muscles dn = 1281:1285 (excluding cases with an initial FAC = 5) BBS, Berg balance scale; FAC, functional ambulation category; FMLL, Fugl-Meyer assessment for lower limb; MBI, modified Barthel index; MMSE, mini-mental state examination; MRC, Medical Research Council scale; RAGT, robot-assisted gait training Table S2 presents detailed baseline comparisons across individual RAGT device. Table S3 summarizes baseline features by initially assigned robot device. The Erigo group exhibited the most severe impairment profile. In contrast, the Exowalk group had the highest baseline function. The Andago and Morning Walk groups occupied intermediate functional levels. Regarding intervention timing, the Exowalk group had the shortest interval from diagnosis to first session (8.8 ± 9.2 days), while the Lokomat group had the longest (23.4 ± 15.2 days) (Fig. 2). Effector rates also varied substantially across devices, ranging from 18.7% in the Erigo group to 84.8% in the Exowalk group (p < 0.001). Device transitions occurred in a minority of patients (n = 176, 13.2% of the RAGT group), who were classified as the Combined group in subsequent analyses. Baseline characteristics across the RAGT group. BBS, Berg balance scale; FAC, functional ambulation category; FMLL, Fugl-Meyer assessment for lower limb; MBI, modified Barthel index; MMSE, mini-mental state examination; MRC, Medical Research Council scale; RAGT, robot-assisted gait training Lokomat patients were stratified into an early group (enrolled before September 2019; n = 224) and a late group (enrolled from September 2019 onward; n = 145), corresponding to the period before and after the full five-device portfolio became available at our institution. The Lokomat-early group had a significantly higher effector rate than the Lokomat-late group (135 [73.4%] vs. 78 [55.3%]; p = 0.001). Baseline functional status was significantly better in the early group, with higher scores on FMLL, MBI, MMSE, and BBS (Table S4). Logistic regression analyses for the effectors The effectors were defined based on the FAC, and Fig. 3 illustrates the changes in FAC levels from baseline to follow-up in both groups. Table 2 shows the proportions of effectors by RAGT device. Exowalk had the highest effector rate (84.8%), whereas Erigo had the lowest (18.7%). Alluvial diagrams showing changes in FAC levels from baseline to follow-up in the RAGT and control groups. Each colored stream represents the transition of patients between FAC levels, with stratum width reflecting the number of patients in each category. Because FAC is an ordinal paired outcome, within-group pre-post changes were evaluated using the Wilcoxon signed-rank test separately for the conventional treatment and RAGT groups. Both groups showed significant improvements in FAC levels from baseline to follow-up (p < 0.001 for all). FAC, functional ambulation category; RAGT, robot-assisted gait training Table 2Proportions of effectors in each Robot-assisted gait training deviceRAGT devicesTotal, naEffectors, n (%)Non-effectors, n (%)Andago172123 (76.4)49 (24.6)Erigo35159 (18.7)292 (81.3)Lokomat349213 (65.5)136 (34.5)Morning Walk144110 (78.0)34 (22.0)Exowalk9678 (84.8)18 (15.2)Combined173114 (67.9)59 (32.1)aExcluding cases with an initial FAC = 5 or missing valuesFAC, functional ambulation category; RAGT, robot-assisted gait training Proportions of effectors in each Robot-assisted gait training device aExcluding cases with an initial FAC = 5 or missing values FAC, functional ambulation category; RAGT, robot-assisted gait training Both groups showed significant improvements in FAC from pre- to post-intervention (p < 0.001 for all). In the overall cohort, younger age (aOR [adjusted odds ratio] = 0.98; 95% CI [confidence interval]: 0.97–0.99; p < 0.001), hemorrhagic stroke type (aOR = 1.50; 95% CI: 1.13–2.00; p = 0.006), and better baseline functional and cognitive status were significant predictors of effector outcomes. Specifically, higher initial FMLL (aOR = 1.09; 95% CI: 1.07–1.11; p < 0.001), MBI (aOR = 1.01; 95% CI: 1.00–1.02; p = 0.028), MMSE (aOR = 1.04; 95% CI: 1.02–1.06; p < 0.001), BBS (aOR = 1.05; 95% CI: 1.03–1.07; p < 0.001), and MRC score for key muscles (aOR = 1.07; 95% CI: 1.04–1.10; p < 0.001) were associated with greater odds of achieving functional improvement. There was no significant association between the RAGT group and effector outcome compared to the control group (Table 3). Table 3Binary logistic regression model for effector prediction in the overall study populationVariablesAdjusted OR95% Confidence intervalp-valueAdjusted GVIFbAge (per year)0.980.97–0.99< 0.0011.26Male0.950.75–1.220.7131.02Primary diagnosis1.08Ischemic strokeReferenceHemorrhagic stroke1.501.13–2.000.006Others1.530.65–3.570.328Hypertension0.750.55–1.010.0621.08Diabetes0.880.68–1.160.3711.06Dyslipidemia1.000.72–1.390.9771.05Atrial fibrillation0.740.53–1.020.0681.09Initial FMLL1.091.07–1.11< 0.0011.48Initial MBI1.011.00–1.020.0281.69Initial MMSE1.041.02–1.06< 0.0011.18Initial FAC0.410.33–0.50< 0.0012.33Initial BBS1.051.03–1.07< 0.0012.12Initial MRC1.071.04–1.10< 0.0011.18Sessions (per session)a1.001.00–1.000.3121.04RAGT group1.120.84–1.490.4481.14aA single session consisted of 30 min of treatmentbNo serious multicollinearity was identifiedBBS, Berg balance scale; FAC, functional ambulation category; FMLL, Fugl-Meyer assessment for lower limb; GVIF, generalized variance inflation factor; MBI, modified Barthel index; MMSE, mini-mental state examination; MRC, Medical Research Council scale; OR, odds ratio; RAGT, robot-assisted gait training Binary logistic regression model for effector prediction in the overall study population aA single session consisted of 30 min of treatment bNo serious multicollinearity was identified BBS, Berg balance scale; FAC, functional ambulation category; FMLL, Fugl-Meyer assessment for lower limb; GVIF, generalized variance inflation factor; MBI, modified Barthel index; MMSE, mini-mental state examination; MRC, Medical Research Council scale; OR, odds ratio; RAGT, robot-assisted gait training Within the RAGT group, similar predictors were observed. Notably, Exowalk users had significantly higher odds of becoming effectors compared to the Andago group (aOR = 2.95; 95% CI: 1.23–7.04; p = 0.015). In contrast, Erigo (aOR = 0.13; 95% CI: 0.06–0.28; p < 0.001) and Lokomat (aOR = 0.47; 95% CI: 0.24–0.91; p = 0.026) were associated with lower odds (Table 4). Subgroup analyses comparing each robotic device to the control group showed that Morning Walk (aOR = 3.25; 95% CI: 1.79–5.87; p < 0.001), and Exowalk (aOR = 4.42; 95% CI: 1.84–10.64, p < 0.001) were significantly associated with a higher likelihood of achieving effector status compared to conventional treatment. In contrast, Erigo was associated with a significantly lower likelihood of effector status (aOR = 0.58; 95% CI: 0.35–0.96, p = 0.033). No statistically significant differences were observed for the Andago, Lokomat, or combined RAGT groups when compared to conventional treatment. Figure 4 presents a forest plot summarizing the aORs and 95% CIs for each robotic device relative to conventional gait training. Full logistic regression model outputs for each robot device are provided in Tables S5–S10. Table 4Binary logistic regression model for effector prediction in the robot-assisted gait training groupVariablesAdjusted OR95% Confidence intervalp-valueAdjusted GVIFcAge (per year)0.980.96–0.99< 0.0011.26Male0.800.58–1.110.1831.04Primary diagnosis1.08Ischemic strokeReferenceHemorrhagic stroke1.521.05–2.200.028Others1.430.43–4.740.556Hypertension0.810.54–1.220.3081.06Diabetes1.000.71–1.420.9791.06Dyslipidemia1.110.74–1.650.6141.06Atrial fibrillation0.750.50–1.140.1751.09Initial FMLL1.061.03–1.08< 0.0011.48Initial MBI1.000.99–1.010.6761.68Initial MMSE1.031.01–1.060.0071.20Initial FAC0.390.28–0.52< 0.0012.42Initial BBS1.041.01–1.070.0032.22Initial MRC1.051.01–1.090.0061.26Sessions (per session)a1.000.99–1.000.1701.09Diagnosis to start (per day)b1.011.00–1.020.3041.08RAGT devices1.13AndagoReferenceErigo0.130.06–0.28< 0.001Lokomat0.470.24–0.910.026Morning Walk1.140.58–2.230.707Exowalk2.951.23–7.040.015Combined0.760.38–1.550.457aA single session consisted of 30 min of treatmentbThe number of days from stroke diagnosis to the first RAGT sessioncNo serious multicollinearity was identifiedBBS, Berg balance scale; FAC, functional ambulation category; FMLL, Fugl-Meyer assessment for lower limb; GVIF, generalized variance inflation factor; MBI, modified Barthel index; MMSE, mini-mental state examination; MRC, Medical Research Council scale; OR, odds ratio; RAGT, robot-assisted gait training Binary logistic regression model for effector prediction in the robot-assisted gait training group aA single session consisted of 30 min of treatment bThe number of days from stroke diagnosis to the first RAGT session cNo serious multicollinearity was identified BBS, Berg balance scale; FAC, functional ambulation category; FMLL, Fugl-Meyer assessment for lower limb; GVIF, generalized variance inflation factor; MBI, modified Barthel index; MMSE, mini-mental state examination; MRC, Medical Research Council scale; OR, odds ratio; RAGT, robot-assisted gait training Forest plot of aORs and 95% CIs for predicting effector status across RAGT devices, using conventional gait training as the reference. Morning Walk and Exowalk were significantly associated with higher odds of achieving effector status, while Erigo showed a lower probability compared to conventional treatment. aOR, adjusted odds ratio; CI, confidence interval; RAGT, robot-assisted gait training When compared with the control group using multivariable logistic regression, neither the Lokomat-early nor the Lokomat-late group demonstrated a statistically significant association with effector status (early: aOR = 1.69; 95% CI: 1.00–2.88; p = 0.052; late: aOR = 1.33; 95% CI: 0.77–2.28; p = 0.303), consistent with the non-significant finding observed for the overall Lokomat group in the primary analysis (Tables S11 and S12). Sensitivity analyses To assess the robustness of our findings, we conducted a sensitivity analysis after excluding patients with extreme values. The proportion of effectors remained comparable between groups (control: 483/1,258 [62.9%]; RAGT: 690/1,257 [58.6%]), with no statistically significant between-group difference (p = 0.064). Multivariable logistic regression models were evaluated as in Table 3 (overall cohort) and Table 4 (RAGT subgroup analysis), and confirmed no significant association between RAGT group assignment and effector status. Device-specific subgroup analyses also yielded results consistent with the primary analysis (Table S13). Tree-based classification A binary CART model was constructed using baseline functional scores to predict effector status. In the RAGT group, the first split was observed at BBS ≥ 2; patients with scores below this threshold had only a 28.2% probability of being effectors. In contrast, those with BBS ≥ 4 or MMSE ≥ 11 had a substantially higher likelihood, up to 76.5% in the terminal node. FMLL on the lesion side and MMSE consistently appeared in deeper splits, highlighting their additive value (Fig. 5). The model demonstrated an accuracy of 74.9%, a sensitivity of 79.6%, a specificity of 68.3%, a PPV of 77.6%, an NPV of 70.8%, and an AUC of 0.771. The difference between training error and cross-validated error was 0.042, close to the cross-validation standard error (xstd = 0.030), indicating limited overfitting within the derivation sample. CART model for predicting effector status using baseline functional scores and RAGT devices. The model achieved an accuracy of 74.9%, a sensitivity of 79.6%, a specificity of 68.3%, and an AUC of 0.771. The first split was determined by BBS, with deeper nodes incorporating FMLL, MMSE, and initial device assignment as secondary predictors. Each node displays the predicted class (effector vs. non-effector) and the proportion of effectors within that subgroup. BBS, Berg balance scale; CART, classification and regression tree; FMLL, Fugl-Meyer assessment for lower limb; MMSE, mini-mental state examination A separate multiclass CART model was used to examine predictors of initial device assignment among effectors. The first split occurred at FAC ≥ 2, with patients below this threshold mostly assigned to Lokomat. Among those with better baseline ambulation (FAC ≥ 4), Exowalk and Morning Walk were more frequently assigned. MMSE and FMLL contributed to secondary splits, suggesting that cognitive and motor severity were key factors in device selection (Fig. 6). As this model was constructed for exploratory purposes rather than clinical prediction, the overall accuracy of 55.1% reflects substantial class imbalance across device groups and should be interpreted accordingly. Multiclass CART model predicting initial robot device assignment among effector patients. This model was constructed as an exploratory analysis to identify the functional thresholds that characterized clinical device selection in practice, rather than as a predictive classification tool. The first split was determined by FAC, with subsequent splits incorporating BBS, FMLL, and MMSE. BBS, Berg balance scale; CART, classification and regression tree; FAC, functional ambulation category; FMLL, Fugl-Meyer assessment for lower limb; MMSE, mini-mental state examination Discussion This retrospective cohort study of 2,947 stroke patients examined the comparative effectiveness of five distinct RAGT devices versus conventional gait training in a real-world rehabilitation setting. The central finding is that the effectiveness of RAGT is not a fixed property of the technology itself, but depends critically on matching the appropriate device to the patient’s baseline functional level. While RAGT did not demonstrate overall superiority to conventional treatment across the unselected cohort, specific devices, particularly those requiring active weight-bearing and participation, were associated with significantly better outcomes in appropriately selected patients. These findings challenge the assumption that RAGT is universally beneficial and support a paradigm shift toward function-guided, individualized device selection. Compared with earlier randomized trials and meta-analyses demonstrating benefits of RAGT under controlled conditions [26–28], our findings suggest that RAGT does not confer a uniform advantage over conventional therapy in unselected, real-world stroke populations. This discrepancy likely reflects differences between efficacy and effectiveness. Nevertheless, our results are consistent with recent evidence indicating that the benefits of RAGT are context-dependent [6, 29]. Importantly, lower baseline FAC scores were associated with a greater likelihood of improvement, suggesting that patients with more severe gait impairment may benefit most from robotic assistance. This aligns with prior evidence that RAGT is particularly effective in non-ambulatory patients and in the early post-stroke phase [2]. The discrepancy between the non-significant overall RAGT effect and the significant device-specific findings likely reflects heterogeneity across devices. Pooling devices with differing effects, including those associated with lower or non-significant odds of improvement, may attenuate the overall effect toward the null [30–32]. Thus, device-specific analyses provide a more clinically meaningful interpretation, highlighting that effectiveness depends on both device type and patient characteristics. Because multiple comparisons were performed without formal adjustment, these device-specific findings should be interpreted as exploratory and hypothesis-generating. Our real-world data provide valuable insights into the clinically tolerable ranges for different RAGT devices based on baseline functional assessments. The CART analysis revealed distinct functional thresholds guiding robotic device selection in practice, with FAC, BBS, FMLL, and MMSE serving as primary determinants. Patients with very poor balance (BBS < 2) had only a 28.2% probability of achieving effector status, while those with preserved balance (BBS ≥ 4) and cognitive function (MMSE ≥ 11) demonstrated markedly higher success rates, reaching up to 76.5%. Device assignment patterns showed that patients with minimal ambulation capacity (FAC < 2) were predominantly allocated to Lokomat, whereas those with better baseline function (FAC ≥ 4) were more frequently assigned to Exowalk or Morning Walk. The functional thresholds identified in our CART model are broadly consistent with emerging FAC-based frameworks for RAGT device selection proposed in the literature. Previous studies proposed a schematic guideline stratifying patients by baseline FAC level to guide the selection of RAGT type and assistance level [33, 34]. Our data-driven thresholds align with this stratification principle. Similarly, the BBS threshold identified in our binary CART model is consistent with clinical observations that patients with severely impaired balance are poor candidates for devices requiring postural self-maintenance [35]. These convergent findings across real-world and guideline-based frameworks strengthen the clinical validity of our CART-derived thresholds. Meanwhile, our results showed that neither the Lokomat-early nor the Lokomat-late group demonstrated a statistically significant association with effector status compared to the control group, consistent with the non-significant finding for the overall Lokomat group in the primary analysis. These findings suggest that the non-significant result for Lokomat is not attributable to temporal confounding from changes in case mix, but rather reflects the genuine clinical characteristics of patients treated with this device across the study period. Among devices showing statistically significant associations in the exploratory subanalyses, Exowalk and Morning Walk demonstrated higher odds of effector status compared to conventional treatment. These findings, if replicated, may be attributed to their shared biomechanical characteristics, which fundamentally differentiate them from other RAGT devices. Both devices eliminate harness-based bodyweight support systems, requiring patients to actively maintain trunk control and balance throughout training. Exowalk functions as an exoskeleton requiring greater active weight bearing and trunk stabilization, whereas Morning Walk offers partial trunk support via a saddle-seat system without overhead suspension [12, 13]. Importantly, Morning Walk, as an end-effector device, differs from exoskeleton systems by not providing robotic leg support during standing and gait phases [36]. Instead, patients must rely on their own lower limb function to perform these tasks, imposing additional motor and postural demands [37]. These challenges are tolerable only for patients with sufficient baseline motor and balance capacity, as evidenced by the preferential assignment of higher-functioning patients to these devices in our cohort. This design requires greater active participation, imposing more demanding motor tasks that may promote motor learning and neuroplasticity through task-specific engagement [38, 39]. The neurobiological rationale aligns with motor learning principles that emphasize active involvement, error-based learning, and appropriate challenge levels for cortical reorganization [40]. Patients able to meet these challenges, those with adequate motor and balance capacity, may benefit from enhanced proprioceptive feedback, improved intermuscular coordination, and more physiological gait patterns compared to passive or highly assistive devices. Therefore, optimizing the challenge level within each robotic device based on individual patient capacity may be critical to maximizing therapeutic benefit. However, the specific role and relative effectiveness of each RAGT device likely vary across institutions depending on available technologies. The advantages observed in our study should be interpreted in the context of our unique combination of robotic systems; institutions with different device portfolios may observe alternate patterns of benefit. Our analysis identified consistent predictors of functional improvement across RAGT devices, including younger age, better baseline motor function, cognitive status, balance, and muscle strength [1]. These factors not only predicted overall treatment success but also influenced the relative effectiveness of specific robotic devices, as evidenced by device-specific effects in our regression analyses. The consistency of these predictors suggests that fundamental neurological and functional capacities serve as universal determinants of recovery potential, regardless of the type of robotic assistance. However, the magnitude and importance of these predictors varied by device, indicating that clinical application must be tailored to individual devices. CART analysis further showed that specific functional thresholds, particularly BBS and MMSE, could effectively predict treatment outcomes and guide device assignment. This supports an evidence-based framework for patient-robot matching [41] that is feasible for implementation in routine clinical settings and underscores the need for comprehensive functional assessment prior to robotic intervention. However, the relatively high sensitivity (79.6%) compared to specificity (68.3%) suggests the model is better suited to identifying patients likely to respond to RAGT than to ruling out non-responders, which should be considered in clinical application. The multiclass CART model, which examined patterns in device assignment among effectors, was constructed as an exploratory analysis rather than a predictive tool. Its overall accuracy of 55.1% reflects substantial class imbalance across device groups and should not be interpreted as evidence of reliable device-assignment prediction; rather, the tree structure illustrates the functional thresholds that characterized device selection in clinical practice at our institution. These findings carry direct implications for clinical practice. The effectiveness of RAGT is not universal but contingent on appropriate patient-device matching guided by baseline functional status, a principle that clinicians should incorporate into standard assessment and prescription protocols. The functional thresholds identified by CART offer a practical, assessment-driven framework for device selection that is feasible within routine inpatient rehabilitation workflows, moving practice away from a one-size-fits-all approach and toward selective, individualized robotic therapy [42, 43]. Several directions for future research emerge from this study. First, prospective multicenter studies with predefined device allocation criteria are needed to validate the CART-identified functional thresholds and establish generalizable patient-robot matching guidelines. Second, head-to-head randomized trials comparing specific RAGT devices within matched patient subgroups would provide more definitive evidence of device-specific effects than the retrospective subgroup analyses employed here. Third, the development and validation of standardized assessment protocols to minimize missing functional scale data, particularly in severely impaired patients, would improve the completeness and quality of real-world RAGT research. Finally, economic evaluations examining the cost-effectiveness of function-guided device selection, relative to both universal RAGT and conventional treatment, would inform resource allocation decisions at the institutional and health-system levels. This study has several limitations. First, its retrospective design limits causal inference, and non-randomized allocation introduces the potential for selection bias, as patients with more severe deficits were preferentially assigned to RAGT, which was reflected in the significantly worse baseline functional scores observed in the RAGT group. Second, although real-world data enhance generalizability, they also reflect confounders such as resource constraints, scheduling limitations, and unmeasured clinical factors that may influence both treatment and outcomes. Additionally, the number of days from diagnosis to intervention initiation was available only for the RAGT group, as this variable was not routinely documented for patients receiving conventional treatment. This precluded a direct comparison of intervention timing between groups, which may be a relevant confounder of recovery outcomes. Third, despite multivariable adjustment, residual confounding from variables such as neuroimaging findings, laboratory parameters, or stroke subtype characteristics likely remains. Fourth, the relatively low cutoff values in the CART analysis (e.g., BBS ≥ 2 out of 56 points) may limit clinical interpretability, raising concerns about practical applicability. Particularly, the models were derived and evaluated within the same single-center dataset without external validation, and the identified functional thresholds, including the low absolute cutoff values, may reflect institution-specific clinical patterns rather than generalizable criteria. Further, the multiclass CART model, which examined device assignment patterns among effectors, showed a more pronounced discrepancy between training and cross-validated error, suggesting moderate overfitting, and achieved a relatively lower accuracy due to substantial class imbalance. Therefore, this model should be interpreted as exploratory and hypothesis-generating rather than as a validated clinical prediction tool. Fifth, missing data in functional scale assessments led to the exclusion of 475 patients from the logistic regression models (24.0% of the primary analysis set of 1,981 patients), suggesting that missingness may be associated with greater clinical severity rather than being entirely random. Sixth, because within-device training parameters were individualized but not uniformly archived across robotic platforms, the present study addresses between-device allocation rather than optimization of device-specific settings. Finally, this single-center study was conducted at a tertiary-level university-affiliated rehabilitation center with access to multiple robotic devices, so findings may not generalize to settings with different patient populations, robotic infrastructure, or clinical workflows and thus require external validation. Conclusion This large-scale retrospective study demonstrates that while RAGT may not outperform conventional gait training for post-stroke ambulation recovery, specific robotic devices can produce superior outcomes when matched to a patient’s functional status. Our key finding, that devices requiring active participation without full bodyweight support are more effective, emphasizes the importance of individualized device selection over universal robotic therapy application. These results support a paradigm shift in robotic rehabilitation: from blanket implementation toward selective, evidence-based patient-robot matching guided by motor, balance, and cognitive profiles. Rather than viewing RAGT as a universally beneficial intervention, our findings advocate for a more nuanced, personalized approach that acknowledges patient heterogeneity and the device-specific potential to address distinct functional deficits and recovery trajectories. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 2. Abbreviations Adjusted odds ratio Area under the receiver operating characteristic curve Berg balance scale Body mass index Body weight support Classification and regression tree Confidence interval Functional ambulation category Fugl-Meyer assessment for lower limb Generalized variance inflation factor Modified Barthel index Mini-Mental State Examination Medical Research Council Negative predictive value Positive predictive value Robot-assisted gait training Author contributions Conceptualization: DYK and DHK; Data Curation: DK and G-MP; Formal Analysis: DP and DK; Investigation: WHC and SYY; Methodology: DP, DK, G-MP, and DHK; Supervision: WHC, YWK, and DYK; Validation: SYY and DYK; Visualization: DP and DK; Writing – Original Draft Preparation: DP and DHK; Writing – Review & Editing: DP, DYK, and DHK; All authors read and approved the final manuscript. Funding This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: RS-2023-00262005). Data availability The datasets used and/or analyzed in the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate The institutional review board of Severance Hospital approved the procedures and protocols of this study on June 17, 2022 (approval no. 4-2022-0546). Owing to the retrospective design, the requirement for informed consent was waived. Consent for publication Not application. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Deog Young Kim, Email: KIMDY@yuhs.ac. Dae Hyun Kim, Email: hohoho7490@gmail.com. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Data Availability Statement The datasets used and/or analyzed in the current study are available from the corresponding author on reasonable request.