Sztuczna inteligencja w fizjoterapii, terapii zajęciowej i neurorehabilitacji: skuteczność kliniczna, wartość prognostyczna i wykonalność wdrożenia – przegląd systematyczny

BACKGROUND: Global rehabilitation needs far exceed capacity, and artificial intelligence (AI) is proposed to extend access, personalise therapy, and support adherence. We aimed to synthesise evidence on clinical effectiveness, prognostic performance, and implementation feasibility of AI-enabled rehabilitation across conditions and care settings. METHODS: We conducted a mixed-methods systematic review of AI-enabled rehabilitation and rehabilitation-led prevention relevant to physiotherapy practice. MEDLINE, Embase, Web of Science, CINAHL, Scopus, and IEEE Xplore were searched. Methodological quality was appraised with MMAT and PROBAST + AI (prediction/diagnostic models), with a priori AI-reporting minimums checklist. Given heterogeneity, evidence was integrated via prespecified thematic synthesis. FINDINGS: Thirty studies from diverse regions met inclusion, comprising randomised trials, non-randomised comparisons, cohorts, surveys, qualitative work, and prediction/diagnostic models. Clinical effects were modest and heterogeneous. Most AI-enabled interventions were comparable to conventional rehabilitation; signals of benefit appeared in selected musculoskeletal and telerehabilitation contexts but rarely persisted beyond short follow-up. Internal validity was frequently limited by asymmetric adherence measurement objective telemetry in AI arms versus self-report or attendance in controls alongside low uptake and declining engagement over time. Safety and usability were generally favourable within short horizons, although surveillance and explicit attribution to AI components were inconsistently reported. Prognostic and adaptive models showed encouraging discrimination in development settings but lacked multicentre external validation, calibration, subgroup-error profiling, and prospective impact evaluation, leaving them unready for clinical use. Stakeholders reported willingness to adopt AI while highlighting gaps in training, governance and information technology support, costs, and digital equity. Overall certainty for comparative clinical outcomes was low to moderate, for models, very low. CONCLUSION: AI in rehabilitation presently acts more as a behavioural amplifier structuring home programmes and supporting execution than as a replacement for dose-matched therapist-delivered care. Credible scale-up should position AI as an adjunct and hinge on arm-symmetric capture of adherence and execution, clinically meaningful outcomes paired with verified behaviour change over treatment and longer-term follow-up, and model deployment only after transparent development, external validation with calibration and subgroup-error characterisation, and demonstration of clinical impact. Embedding equity and economic evaluation, together with living oversight and version control, will be essential to convert promising prototypes into trustworthy, durable, and widely accessible services.