Use of artificial intelligence for outcome assessment in pediatric rehabilitation: a scoping review

Abstract

Artificial intelligence (AI) is increasingly being applied in healthcare, with growing potential to enhance rehabilitation. In pediatric rehabilitation, traditional outcome measures are resource-intensive, time-consuming, and prone to variability, limiting their scalability. AI offers opportunities to automate, standardize, and expand access to outcome assessment. However, the scope, methodological rigor, and clinical utility of AI applications in this field remain unclear. This scoping review examined how AI has been applied to pediatric rehabilitation outcome assessment, focusing on populations studied, AI methods and models applied, outcome domains, stage of implementation, and reported limitations. A scoping review was conducted in accordance with PRISMA-ScR guidelines MEDLINE (Ovid), CINAHL (EBSCOhost), Embase (Ovid), and IEEE Xplore were searched from database inception to June 9, 2025, yielding 11,370 records; 51 studies met the eligibility criteria and were included. Study selection followed the Population, Concept, and Context (PCC) framework. Screening and data extraction were performed in Covidence by three reviewers with piloting at each stage. Data were synthesized descriptively in tables and narrative summaries. Reported AI model performance was extracted as the highest metric provided in each study (≥ 90% accuracy, F1-score, sensitivity, or specificity). Fifty-one studies met the inclusion criteria. Most studies were exploratory and conducted at preclinical or early pilot stages. Children with cerebral palsy were the most frequently studied population, particularly in relation to gait analysis. AI applications were predominantly focused on motor-related outcomes, including gait, movement quality, upper-limb function, and ambulation ability, while non-motor domains such as cognitive or behavioral outcomes were sparsely represented. Supervised machine learning was the most commonly used AI type, followed by neural networks and deep learning approaches, with model selection closely aligned to data modality and task requirements. AI was most often applied for classification, prediction, and automated quantification or scoring. While several studies reported high performance for specific tasks, methodological heterogeneity, limited external validation, and small sample sizes constrained comparability and clinical translation. AI-based outcome assessment in pediatric rehabilitation is an emerging and rapidly evolving field, with the strongest evidence to date in motor-related applications, particularly gait analysis. Current AI tools remain largely supportive and analytical, rather than integrated into real-time clinical decision-making. Future research should prioritize methodological rigor, broader representation of pediatric populations and outcome domains, feasibility and implementation studies, and explicit consideration of ethical and equity-related issues to support responsible and clinically meaningful adoption of AI in pediatric rehabilitation.

Ähnliche Arbeiten