How Accurately Can Artificial Intelligence Measure Component Position After Total Hip and Knee Arthroplasty? A Systematic Review and Meta-Analysis

Background/Objectives: Artificial intelligence (AI) may standardise radiographic measurement of component position, but its accuracy is uncertain. We assessed agreement between AI and reference-standard measurements after primary total hip (THA) and total knee arthroplasty (TKA). Methods: Four databases were searched from 2015 to 13 July 2026 (PROSPERO CRD420261455545). Study selection, data extraction, risk of bias (QUADAS-2), reporting (CLAIM 2024), and certainty of evidence (GRADE) were assessed in duplicate. Signed mean differences were pooled with random-effects models and modified Hartung–Knapp intervals, while proportions within 2° were pooled using binomial-normal mixed models. Results: Nineteen studies were included, and six contributed to meta-analysis. For THA, pooled directional mean differences were +0.05° (95% CI −0.27 to +0.37) for cup anteversion and −0.13° (95% CI −0.89 to +0.63) for inclination; reported study-level mean absolute errors reached 3.59° and 1.79°, respectively. For TKA, the study-level agreement within 2° ranged from 70.4% to 99.3%; pooled values ranged from 88.4% to 95.1%, with broad confidence and prediction intervals. Ten studies were at high risk of bias, and external testing was confirmed in only five. Certainty of evidence was very low for all five outcomes. Conclusions: Pooled directional mean differences for cup orientation were close to zero, but this does not establish accurate individual measurements. Knee estimates were heterogeneous and imprecise. Current evidence does not establish reliable autonomous clinical use; clinically important or borderline values require expert assessment.