From data to decisions: machine learning in predicting outcomes of robotic-assisted total knee arthroplasty

Machine learning (ML) has emerged as a transformative tool for outcome prediction in total knee arthroplasty (TKA). The integration of ML with robotic-assisted TKA (RA-TKA) systems offers potential advantages in surgical planning and personalized patient care. This systematic review synthesizes current evidence on ML applications for predicting outcomes in RA-TKA patients.

A systematic literature search was conducted using PubMed, Embase, Web of Science, and Cochrane Library for studies published through January 2025. Two reviewers independently screened articles and assessed methodological quality using PROBAST and MINORS tools. Studies evaluating ML algorithms for predicting surgical outcomes in TKA patients were included, with particular attention to RA-TKA applications.

Forty-seven studies met inclusion criteria. ML models demonstrated superior performance over traditional risk scores for predicting periprosthetic joint infection (AUC 0.82-0.88), venous thromboembolism (AUC 0.85-0.89), and transfusion requirements (AUC 0.83-0.86). Notably, simple three-variable models (age, sex, ASA) explained 70% of mortality variation, challenging the necessity of complex algorithms. RA-TKA showed modest clinical benefits over conventional techniques, though cost-effectiveness remains unproven.

ML has transitioned from experimental to clinical application in RA-TKA outcome prediction, with models demonstrating good-to-excellent discriminative ability (AUC 0.82-0.89). However, this review reveals critical tensions that warrant attention: simple models with few variables often perform comparably to complex algorithms, challenging the assumption that sophistication equates to clinical utility; only a minority of included studies specifically addressed RA-TKA populations (25.5%), limiting the generalizability of current prediction models to robotic surgery cohorts; and no validated prediction model currently incorporates real-time intraoperative data from robotic systems. Critical challenges include data quality, algorithm interpretability, validation across diverse populations, and the need for cost-effectiveness analyses. Future research should prioritize head-to-head model comparisons, RA-TKA-specific model development, and strategies to mitigate algorithmic bias.