Artificial Intelligence in Arthroplasty: A Comprehensive Structured Critical Review and Descriptive Evidence Map of Validation, Uncertainty, and Workflow Integration

Background/Objectives: Artificial intelligence (AI) in arthroplasty spans clinical prediction, imaging, implant identification, planning, natural language processing, infection support, patient communication, and workflow optimization, but its translational readiness remains uncertain. This review evaluated how often current studies move beyond internal model development toward credible external evaluation, uncertainty-aware reporting, and clinically meaningful integration. Methods: A structured critical narrative review and descriptive evidence map included 252 publications identified and verified between 1 January and 17 August 2026. Of these, 172 were core arthroplasty-specific primary-AI publications used for task mapping, whereas 150 involved an assessable fitted model or operational system suitable for transportability analysis; 139 publications belonged to both analytical sets. Transportability was assessed separately based on geographic/site independence, temporal separation, and model state at the time of evaluation. Results: Strict frozen-model external evaluation was identified in 27/150 publications (18.0%). Twelve of these 27 publications (44.4%) belonged to a single commercial planning program; treating those 12 publications as a single program-level evidence unit reduced the count from 27 to 16. Later, non-overlapping temporal evaluation was established in 7/150 publications (4.7%), while the temporal relationship remained unclear in 86/150 (57.3%). In eight of the 150 transport-assessable publications (5.3%), external-validation terminology mapped to a different operational category under the framework used here. Among the 172 core publications, confirmed-present lower bounds were 27/172 (15.7%) for calibration, 5/172 (2.9%) for uncertainty or out-of-distribution handling, and 8/172 (4.7%) for observed workflow or patient benefit. The strongest near-term evidence concerned bounded, auditable tasks such as implant recognition, measurement, planning, and structured information extraction. Conclusions: Arthroplasty AI is best positioned as clinician-supervised precision-support infrastructure. Credible translation requires clearly defined evaluation settings, transparent reporting of model state, calibration and uncertainty assessment where applicable, and prospective workflow-level evaluation.