# Artificial intelligence in rehabilitation: a review of clinical effectiveness, real-world performance, safety, and equity across modalities and settings

## TL;DR

AI and devices like robotics can help more people get personalized rehabilitation, but their real-world effectiveness, safety, and fairness need more rigorous testing and monitoring.

## Contribution

The paper introduces a framework for evaluating AI in rehabilitation, emphasizing clinical standards, equity, and post-market monitoring.

## Key findings

- Technology-assisted training improves activity outcomes for post-stroke upper limb but lacks consistent effects on impairment and independence.
- AI-enabled systems show a performance drop from development to deployment, especially in brain-computer interfaces and movement evaluation.
- Adverse events are mild, but usability, equity, and cost in home-based delivery remain under-researched.

## Abstract

Rehabilitation faces a scale problem: millions who could benefit lack timely, effective services. Artificial intelligence (AI) and device-based modalities (e.g., robotics and VR) can extend reach and personalise care when validated, yet decision-makers lack a consolidated view of clinical usefulness, translation to practice, safety, equity, and cost.

We conducted an umbrella review of reviews using a Population–Exposure–Outcome framework. Searches span biomedical, allied health, and engineering databases from inception to September 1, 2025. We distinguished AI-enabled (ML/DL) interventions from technology-assisted (no ML demonstrated) modalities and synthesised outcomes across impairment, activity, independence, usability/safety, equity, and economics.

The most reproducible clinical signal is activity improvement for post-stroke upper limb with technology-assisted training (robotics with or without VR) that increases task-specific practice; effects on impairment and independence are inconsistent once dose is matched and assessors are blinded. Claims of non-inferiority are not established when prespecified margins and confidence-interval testing are absent, so parity is interpreted as no between-group advantage under those conditions. Across AI-enabled domains, a development-to-deployment performance drop is evident most notably for brain–computer-interface classifiers and computer-vision movement evaluation limiting immediate clinical impact. Imaging-based decision support (radiomics/CNN) is closer to practice but varies by software and site, requiring local calibration and impact evaluation before pathway change. Reported adverse events are generally mild, yet usability, adherence, equity, and cost are under-measured, particularly in home and hybrid delivery. Prediction-model and trial reporting frequently fall short of contemporary AI standards; representation skews toward high-income settings, and subgroup performance is seldom reported.

An adjunct-first posture is warranted. Adoption should be gated by minimum clinically important difference–anchored benefit under dose symmetry and blinded assessment; external, multi-site validation with declared lab-to-clinic performance loss; subgroup fairness with mitigation; decision-grade economic value; interoperability; and readiness for regulation, change control, and cybersecurity. Priorities include pragmatic, multi-site, assessor-blinded, dose-matched trials; standardised safety/usability capture for home use; and a public, living evidence atlas. AI can expand rehabilitation when held to clinical standards that matter to patients and services. With clear adoption gates and continuous post-market monitoring, systems can extend access and independence without sacrificing rigour, safety, equity, or fairness.