# The effects of suspension-supported training on dynamic balance capacity in stroke patients: a systematic review and meta-analysis enhanced by XGBoost machine learning

## TL;DR

Suspension-supported training improves dynamic balance in stroke patients, with optimal training duration and frequency suggested based on machine learning analysis.

## Contribution

This study integrates XGBoost machine learning with a meta-analysis to explore the optimal training dose for dynamic balance improvement in stroke patients.

## Key findings

- SST significantly improves dynamic balance in stroke patients with a moderate effect size.
- Training duration and intervention mode were identified as the most influential predictors of outcome.
- A regimen of 90–100 minutes per session, 6 times per week for 8–9 weeks is suggested for favorable results.

## Abstract

Systematically evaluate the effectiveness of suspension-supported training (SST) on dynamic balance in stroke patients and determine its optimal training dose to establish an evidence-based foundation for clinical rehabilitation.

Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, randomized controlled trials (RCTs) meeting Population, Intervention, Comparator, Outcome, Study design (PICOS) criteria were screened in PubMed, Web of Science, PsycINFO, and the Cochrane Library (search cutoff: 31 October 2025). Twelve studies comprising 584 patients were included. A restricted maximum likelihood (REML) random-effects model was used to pool standardized mean differences (SMDs). Heterogeneity was assessed using the I-squared (I2) statistic and Cochran’s Q test. Publication bias and robustness were evaluated by sensitivity analysis, Egger’s regression test, and the trim-and-fill method. Subgroup analyses and meta-regression were conducted to explore potential moderators; additionally, a Bayesian dose–response meta-analysis was performed to characterize the non-linear relationship between training dosage and dynamic balance outcomes. An exploratory extreme gradient boosting (XGBoost) model combined with SHapley Additive exPlanations (SHAP) analysis was further applied to investigate dose–response patterns and feature importance.

SST produced a significant improvement in dynamic balance (SMD = 0.87, 95% CI 0.49–1.26, p < 0.0001), with substantial heterogeneity across studies (I2 = 76.6%). The overall quality of evidence was rated as moderate using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach. Larger effects were observed in interventions lasting 6–8 weeks, sessions of 40–150 min, patients in the subacute phase, combined intervention approaches, and protocols employing approximately 30% body-weight support. Exploratory XGBoost–SHAP analysis identified training duration and intervention mode as the most influential predictors of outcome. Preliminary dose–response modeling suggested that a regimen of ~90–100 min per session, six sessions per week for 8–9 weeks (total dose ≈ 34.6–37.6 h; ~36 h) might be associated with favorable dynamic-balance gains; however, these estimates should be regarded as hypothesis-generating.

SST effectively enhances dynamic balance in stroke patients. Preliminary models indicated that a regimen of 90–100 min per session, 6 times per week for 8–9 weeks (total duration: 34.6–37.6 h, ~36 h) might yield favorable results. However, given the heterogeneity and study-level limitations, these dosage findings were considered hypothesis-generating. Future research should prioritize high-quality, large-sample RCTs with long follow-up to verify clinical and long-term effectiveness.