Robot-assisted gait training in people with subacute and chronic stroke: a randomized controlled trial — complement, not replacement

Clinical Question (PICO)

In adults with a first ischemic or hemorrhagic stroke, gait impairment, and capacity to participate in inpatient rehabilitation, does four weeks of end-effector robot-assisted gait training plus conventional physiotherapy (RAGT+CPT), compared with time-matched conventional physiotherapy alone (CPT), improve walking independence and basic activities of daily living? Stroke phase was a prespecified stratification factor: subacute, approximately 7 days through 6 months, versus chronic, more than 6 months. This September 2026 trial is reviewed here in its original inpatient context, not as a statement of current guideline recommendations. Trial report.

Bottom Line

RAGT+CPT did not demonstrate clear superiority over CPT for either primary outcome: the treatment-by-time interaction was OR 0.60 (95% CI 0.12–2.90; P=0.524) for walking category and OR 0.22 (95% CI 0.03–1.82; P=0.160) for Barthel dependency category. Both groups improved; better self-perceived health favored RAGT among secondary outcomes, but these findings support a complementary role rather than universal superiority—and do not establish equivalence. Trial results and conclusion.

Design

Trial type: Single-center, parallel-group, assessor-blinded randomized superiority trial; participants and treating therapists could not be blinded to the device.
N: 100 randomized, 50 per intervention.
- RAGT+CPT: 25 subacute and 25 chronic.
- CPT: 25 subacute and 25 chronic.
Randomization: 1:1, stratified by stroke phase, with permuted blocks of concealed size; an independent researcher prepared opaque, sequentially numbered sealed envelopes.
Setting: Clinic for Rehabilitation Münster, Austria; multidisciplinary inpatient rehabilitation.
Enrollment: October 18, 2022–March 17, 2025.
Follow-up: Four-week intervention and eight additional weeks, to Week 12. Primary FAC and BI measurements were baseline and Week 4 only; Week 12 follow-up chiefly covered questionnaire outcomes and falls, not a repeat full objective walking battery.
Analysis: Intention-to-treat as reported. Cumulative link mixed models for collapsed ordinal scales; mixed ANOVA for continuous outcomes. Isolated missing observations were addressed using multiple imputation under a missing-at-random assumption. Additional baseline-FAC-adjusted continuous-outcome mixed models were post hoc.
Primary outcomes: Functional Ambulation Categories (FAC, 0–5), analyzed as 0–3, 4, or 5; and Barthel Index (BI, 0–100), analyzed as 0–60, 61–90, or 91–100. One FAC level was considered clinically meaningful; no BI minimal clinically important difference was applied.

Population

Inclusion Criteria

Age at least 18, first-ever ischemic or hemorrhagic stroke, and neurologist-confirmed gait impairment with FAC 0–4.
Subacute or chronic stroke, with no upper time limit for the chronic stratum.
German fluency and sufficient comprehension to follow procedures; Mini Mental Status Test at least 21/30.
Nonambulatory participants could enter if they could sit upright unsupported, with feet supported, for at least one hour twice daily.

Exclusion Criteria

Medical, orthopedic, vascular, neurological, or psychiatric conditions compromising safe active rehabilitation or interfering with treatment.
Relevant skin, bone, lower-limb sensorimotor, spasticity, cardiovascular, or seizure-related contraindications; pregnancy; recent medication changes affecting performance; or failure to meet device-specific height or weight limits. The main paper refers to Supplementary File S1 for the complete operational list.

Baseline Characteristics

Mean age 67.90±11.72 years; 62 men and 38 women; 83 ischemic and 17 hemorrhagic strokes.
Mean time since stroke: 0.27±0.17 years in the subacute stratum and 8.07±13.27 years in the chronic stratum.
Median NIHSS across the four cells: 7, 6, 8, and 6; median BI: 90, 85, 80, and 80, respectively, in the order subacute RAGT, subacute CPT, chronic RAGT, chronic CPT.
Fifty-two participants already had FAC 4: independent on level surfaces, though not necessarily on all surfaces. Severe gait impairment was unevenly distributed: seven participants with FAC 0 were in RAGT arms and none in CPT arms.

Interventions

RAGT+CPT: Lexo end-effector gait trainer with dynamic body-weight support and pelvic guidance. The methods describe 45–50-minute robot sessions four times weekly for four weeks, plus 25-minute CPT sessions twice weekly. Results report a mean of 15 robot sessions (median 16, range 12–16) plus eight CPT sessions. Time included setup, transfers, safety checks, rest, and active practice.
CPT: Individualized therapist-supervised overground walking, stairs, transfers, and balance training. Results specify sixteen 45–50-minute and eight 25-minute CPT sessions, intended to match the combined arm's scheduled time.
Both: Progressive, tolerable task practice and routine multidisciplinary inpatient care. Step counts were not matched, and perceived exertion was monitored rather than used to equalize dose. The abstract and methods compress the weekly schedule differently; the completed-session description above is the clearest reproducible account, not evidence that active stepping time was identical.

Outcomes

Primary Outcomes at Week 4

The following are time-by-treatment interaction odds ratios, representing the CPT change relative to the RAGT+CPT change in odds of a better category. They are not simple post-treatment odds ratios or risk ratios.

Outcome	Interaction OR (95% CI)	P value	Interpretation
Higher FAC category	0.60 (0.12–2.90)	0.524	No demonstrated differential change
Better BI dependency category	0.22 (0.03–1.82)	0.160	No demonstrated differential change

Within the RAGT reference group, the post-versus-baseline OR was 42.40 (6.57–273.57; P<0.001) for FAC and 28.54 (2.90–281.08; P=0.004) for BI. These large within-group changes are not the randomized comparison. Across the cohort, dependent/supervised walking (FAC 0–3) decreased from 48 to 25 participants and FAC 5 increased from zero to 30. Primary results.

Stroke-phase strata and exploratory responses

At least one FAC-level improvement: subacute RAGT 12/25 (48%) versus CPT 8/25 (32%); chronic RAGT 8/25 (32%) versus CPT 11/25 (44%). These are descriptive arm-level proportions, not confirmatory subgroup effects with reported confidence intervals.
Phase stratification was prespecified, but stable phase-specific primary treatment-effect ORs and confidence intervals were not reported. The ordinal models adjusted for phase without the three-way interaction because more complex models were unstable. A significant response in one cell cannot establish a treatment-by-phase interaction.

Secondary Outcomes

Walking speed: No overall time-by-treatment interaction (P=0.629) or three-way interaction (P=0.590). An exploratory post-hoc subacute contrast favored RAGT by 0.234 m/s (95% CI 0.007–0.461; P=0.043); the chronic contrast was reported as −0.096 m/s (−0.352–0.133; P=0.405). These contrasts should not override the negative interaction tests. The apparent time-by-phase effect was no longer significant after baseline-FAC adjustment (P=0.537).
Walking endurance: The 6-minute walk three-way interaction was significant (P=0.036), remaining so after post-hoc baseline-FAC adjustment (P=0.046). This suggests a phase-dependent pattern, not a proven robot advantage within every phase. Table 2 reports a subacute RAGT-versus-CPT contrast of 41.03 m (95% CI −80.44–162.51). The narrative and table differ for some related contrasts and P values; do not treat any one borderline comparison as definitive.
Functional mobility: No differential TUG change (time-by-treatment P=0.482); overall RMI time-by-treatment interaction P=0.316.
Self-perceived health: EQ-5D-5L VAS favored RAGT in the ordinal model: interaction OR 0.23 (0.05–0.98; P=0.047) at Week 4 and 0.05 (0.01–0.23; P<0.001) at Week 12, with overall interaction P=0.001. As above, these encode the smaller CPT change. EQ-5D index and SF-SIS did not show corresponding differential change.
Fatigue: A Week 4 interaction favored RAGT (OR 17.11, 95% CI 1.59–183.59; P=0.019, with higher categories indicating worse fatigue), but the Week 12 interaction was not significant (P=0.117), and the overall interaction was P=0.051.
Falls: Participants reporting at least one fall decreased from 35/100 at baseline to 7/100 at Week 4 and 7/100 at follow-up, without significant between-group differences. Baseline recall and subsequent prospective reporting covered different observation periods; this is not a robot-specific fall-prevention estimate.

Adverse Events / Safety

Thirty-four adverse events occurred in 31 participants. One serious event, a fall-related femoral-neck fracture, was judged unrelated to treatment. Three mild events were considered possibly related to RAGT; there were no severe intervention-related events. Seven participants withdrew from subacute RAGT, compared with one in each other cell; the report identifies two possibly fatigue-related withdrawals in subacute RAGT and one possibly knee-pain-related withdrawal in chronic RAGT. The small trial cannot exclude uncommon harms. Safety and participant flow.

Figures

Participant flow, stratified by stroke phase — *Figure 1.* CONSORT flow diagram of this study. Participants were stratified by stroke phase (subacute vs. chronic) prior to randomization to robot-assisted gait training plus conventional physiotherapy (RAGT+CPT) or conventional physiotherapy alone (CPT).

Lexo end-effector gait trainer; original device photographs — *Figure 2.* Lexo robotic gait trainer (Tyromotion) used for this study. The images show a patient using a harness suspension system for weight support, with connections at the back and pelvis providing trunk support and pelvic guidance. The highly adaptable hardware and software facilitate demanding self-activity during specific gait training. These highlighted areas: (A) Patient unit; (B) Gait unit: sensor-equipped adjustable footplates and foot fixation; (C) User interface with augmented performance feedback. © Tyromotion GmbH.

Walking independence and Barthel Index before and after rehabilitation — *Figure 3.* Changes in walking ability and basic activities of daily living from baseline to post-intervention by treatment group and stroke stage. (A) Functional Ambulation Categories (FAC) and (B) Barthel Index (BI) scores are shown for participants with subacute and chronic stroke, separately for RAGT+CPT and CPT. In the paired plots, blue points and bars represent baseline, and red points and bars represent post-intervention. Individual points represent participant-level observations, bars represent group summary values, and gray lines connect repeated measurements within participants. Statistical analyses were performed using cumulative link mixed models (CLMMs) based on collapsed FAC and BI categories, with fixed effects for time, treatment group, the Time × Treatment interaction, and stroke stage, and a participant-specific random intercept. Brackets indicate model-based within-group comparisons between baseline and post-intervention within each treatment group and stroke-stage stratum. Higher FAC scores indicate greater walking independence, and higher BI scores indicate greater independence in activities of daily living. Significance levels are indicated as *p < 0.05; **p < 0.01; ***p < 0.001. FAC, Functional Ambulation Categories; BI, Barthel Index; RAGT+CPT, robot-assisted gait training plus conventional physiotherapy; CPT, conventional physiotherapy.

Individual Functional Ambulation Categories before and after rehabilitation — *Figure 4.* Individual changes in walking ability across groups. The scatter dot plot displays individual participants’ walking ability at baseline (Week 0, blue dots) and post-intervention (Week 4, ochre dots), assessed using the Functional Ambulation Categories (FAC), where higher scores indicate better walking function. Numbers within the dots indicate the number of participants with each FAC score. Vertical bars represent the median; ns, not significant; *p < 0.05; **p < 0.01; ***p < 0.001. RAGT, robot-assisted gait training; BI, baseline; CPT, conventional physiotherapy; PI, post-intervention.

Criticisms

No superiority is not equivalence. The primary confidence intervals are wide, and no noninferiority margin was tested.
Ceiling effects and imbalance: 52% already walked independently on level surfaces; the robot arms included more severely impaired walkers. Continuous-outcome sensitivity analyses help, but baseline-FAC-adjusted ordinal models were unstable and not interpreted.
Subgroups and multiplicity: Four cells of 25 participants, numerous secondary endpoints, post-hoc contrasts, and simplified ordinal models make selective emphasis on favorable P values risky.
Attrition and missingness: Ten withdrawals were unevenly distributed. Although ITT and imputation are reported, the primary ordinal models contain 190 rather than 200 repeated observations. Their numerical results should not be assumed to remove all potential attrition bias.
Measurement and reporting: Coding inability to perform TUG as zero is clinically distinct from a true zero-second completion; interpretation depends on analytic handling. The paper's narrative and Table 2 also contain inconsistencies in walking contrasts and P values, which limit precision of secondary interpretation.
Dose and durability: Scheduled time is not active step dose; comparative staff workload and cost-effectiveness were not established. Primary walking and ADL outcomes were not objectively reassessed at Week 12.
Causal scope: Both arms received intensive rehabilitation. Without a no-rehabilitation group, within-arm improvements do not isolate rehabilitation from natural recovery or other care. German-speaking, cognitively able patients at one Austrian center do not represent all stroke survivors.

Funding

Tyromotion GmbH supplied the Lexo gait trainer to the Clinic for Rehabilitation Münster. The authors report no financial or other donations to study personnel. The sponsor fulfilled European Medical Device Regulation responsibilities but reportedly had no role in design, collection or analysis of data, the publication decision, or manuscript preparation. The authors declared no commercial or financial conflict of interest. Registration: DRKS00029910, prospectively registered October 17, 2022; amendments approved May 25, 2023. Funding and methods.

The paper

Authors. Michaela Stampfer-Kountchev et al.
Title. Robot-assisted gait training in people with subacute and chronic stroke: a randomized controlled trial.
Journal. Frontiers in Neurology.
Year. 2026; published September 3, 2026.
DOI. 10.3389/fneur.2026.1898475
PMCID. PMC13581499
PMID. 42755508

Deep Dive — click to expand

What this is

This is a test of whether a particular robotic training program adds benefit when compared with a serious, time-matched conventional rehabilitation program—not a test of rehabilitation against inactivity. In 100 inpatients, neither primary interaction established superiority. The September 2026 report is useful because it makes the comparator visible: a device has to outperform the therapy time it displaces, not merely produce impressive before-and-after gains. The commentary below is editorial interpretation, not a guideline or individual treatment recommendation.

1. Shadow Audit

The seductive number is the FAC within-RAGT OR of 42.40. The decision-relevant number is the randomized change interaction, OR 0.60 with a confidence interval spanning 0.12–2.90. Confusing those answers converts recovery during treatment into proof of recovery because of the robot.

A second shadow is the apparently targeted subacute benefit. Twelve subacute robot participants versus eight conventional participants improved at least one FAC level, but the chronic cells move the other way: eight versus eleven. Pooling these descriptive counts gives 20/50 versus 19/50—not a trial-derived adjusted effect estimate, and not evidence of a validated responder subgroup. Meanwhile, the robot arms contained all seven baseline FAC-0 participants. The story is heterogeneity and uncertainty, not a simple winner. Underlying trial data.

2. Inversion Engine

For a claim of universal robot superiority to hold, the randomized primary comparison would need persuasive separation rather than a large within-arm time effect. The FAC interaction's null value is 1; its 0.12–2.90 interval crosses that null widely. For the opposite claim—robots add nothing—to hold, the trial would need enough precision to exclude a clinically important benefit. It does not.

The subacute speed contrast is another stress test: its lower confidence limit is only 0.007 m/s, while the treatment-by-phase-by-time test is P=0.590. The editorial inversion is not that a tiny shift in P proves failure; it is that a favorable subgroup contrast cannot establish that stroke phase changes the treatment effect when the appropriate interaction does not support it.

3. Second-Order Catalyst

The first institution likely to change its process is an inpatient service that already owns a gait robot and must allocate finite therapist and machine time. A reasonable evidence-informed evaluation would retain the device as one option, record baseline FAC and stroke phase, and audit active stepping, functional independence, fatigue, staffing, and patient-valued outcomes against a strong conventional pathway. That is an implementation hypothesis—not an outcome this trial demonstrated.

The patients who most urgently need better evidence are severely dependent walkers and those in the earliest recovery period. They were not the dominant population here. This study cannot establish that they should be prioritized for, or excluded from, the device.

4. Asymmetric Leverage

The asymmetric opportunity is avoiding expensive substitution without evidence of incremental patient benefit. With 52 participants already at FAC 4, buying more capacity on the assumption of universal superiority is a larger extrapolation than offering a monitored complementary option. Conversely, abandoning robotics because of one negative trial would ignore both imprecision and the wider evidence.

The practical payoff might be comparable clinical outcomes with less physical therapist burden or more repetitions—but neither staffing savings nor cost-effectiveness was established. Those are worthwhile targets for a service evaluation, not benefits to insert into a purchasing justification as facts.

5. Paradigm Destroyer

The habit this paper challenges is treating advanced equipment as an intrinsically higher-intensity or more effective treatment. Here, both strategies were embedded in substantial supervised practice, and scheduled time did not establish equal active step dose.

A proposed two-sentence local pathway update: “Offer robot-assisted gait training as a complementary modality selected for a documented rehabilitation goal, rather than as a presumed superior replacement for conventional therapy. Record functional change, tolerance, and actual training exposure, and do not infer treatment success solely from device-session progress or an uncontrolled before-and-after result.” This is a quality-improvement proposal for local review, not a new professional guideline.

MVP — Minimum Viable Proof

The minimum convincing next result would be a replicated randomized advantage in a patient-important functional endpoint, with a confidence interval excluding no benefit, in a clearly defined population and against comparable active practice. A one-level FAC improvement is the trial's clinically meaningful anchor, but demonstrating an extra proportion of patients achieving it requires a prespecified responder analysis—not attaching that threshold to these category-model odds ratios. This trial does not supply a sufficiently precise incremental benefit to mandate routine replacement of CPT.

Best Combination

Read this trial alongside, not instead of, the 2025 Cochrane review. That review included 101 studies and 4,224 participants; the walking-independence analysis involved 51 studies and 2,148 participants and favored electromechanical training plus physiotherapy (OR 1.65, 95% CI 1.21–2.25; moderate-certainty evidence). Its search ended in December 2023, so it did not incorporate this 2026 report.

The synthesis is neither “the review proves every robot works” nor “this trial overturns the review.” Device type, population, comparator, and treatment exposure vary. Use the pooled evidence to establish a plausible class-level benefit, and this trial to question how much additional benefit to expect from this specific substitution in a predominantly ambulatory inpatient population. The index paper's reference list misidentifies the version suffix for the 2025 review; the link above uses the independently verified pub6 record rather than reproducing that bibliographic error.

Overvalue Warning

Do not equate nonsignificance with interchangeability. Wide confidence intervals, uneven withdrawal, and no noninferiority design leave clinically important differences unresolved.
Do not promote secondary or subjective signals into universal practice claims. A favorable EQ-5D VAS signal and exploratory subacute speed contrast are interesting, but they do not establish lasting objective walking superiority, a validated selection rule, or cost savings.

Educational use notice. The Time Is Brain Review is dedicated to medical education. Figures from published articles are embedded for teaching purposes, with attribution to the original publication. We do not claim copyright over any embedded figure; all rights remain with the original publisher and authors under the article's published license. If you are a rights holder and would like a figure removed, please contact the editorial team. See our editorial notice for the full policy.