AI-BASED SURGEON AND PATIENT-SPECIFIC INTRAOPERATIVE PLANNING ALGORITHM FOR TOTAL KNEE ARTHROPLASTY

Background Total knee arthroplasty (TKA) is one of the most commonly performed orthopedic procedures worldwide, with outcomes critically dependent on accurate implant sizing and placement across 14 degrees of freedom. Despite advances in robotic-assisted systems — such as the Smith & Nephew CORI Surgical System with its intraoperative digital tensioner — current rule-based planning algorithms do not account for individual surgeon preferences or patient-specific anatomy. This forces surgeons to make frequent and time-consuming manual adjustments intraoperatively, a recognized source of surgical variability and extended operative time. Prior AI work has focused on preoperative sizing prediction from demographic or imaging data (accuracy within ±1 size: 88–99%), yet no system has addressed real-time, surgeon-personalized intraoperative planning using ligament gap data. Methods We developed and evaluated three machine learning configurations using retrospective data from 7,638 TKA procedures performed by 75 surgeons on the CORI system (August 2020 – July 2024). Each case included native, derived, and purpose-built architected features, including intraoperative gap measurements — available uniquely through the CORI digital tensioner. We independently predicted all 14 planning parameters using Ridge regression with Recursive Feature Elimination (RFE) or Light Gradient Boosting Machine (LightGBM), selected per-parameter via 10-fold cross-validated hyperparameter tuning. Configuration 1 encoded surgeon identity directly; Configurations 2 and 3 applied K-means clustering (K=3) over planned implant placement values and planned gap values, respectively, to assign surgeons to interpretable planning groups — an approach offering regulatory and logistical advantages for AI-enabled medical devices. Results Configuration 1 achieved a mean R² of 0.717 (RMSE = 1.01 mm/°) across all 14 parameters (Figure 1). Compared to the CORI rule-based baseline, AI-generated plans reduced required surgeon adjustments by a mean of 52.8% (20.0 adjustments saved; range: 16.6%–66.7%). Clustering configurations performed comparably: Configuration 2 yielded 44.5% savings (R² = 0.67) and Configuration 3 yielded 43.6% savings (R² = 0.67), with all 75 surgeons experiencing measurable improvement in every configuration (Figure 2). Surgeons who made the most adjustments without AI assistance showed the greatest benefit. Conclusion An AI-driven intraoperative planning algorithm trained on real-world CORI data reduces plan adjustments by over 50% compared to rule-based approaches, with practical clustering configurations offering a viable regulatory pathway. Integration of intraoperative ligament gap data as a predictive feature is a distinguishing and clinically meaningful contribution. Future directions include outcome-weighted training, post-processing guardrails, and surgeon-specific models as data scales.