Advancing preoperative planning technology in total joint arthroplasty with a real-time machine learning calculator:...

Technological innovation in total joint arthroplasty (TJA) has largely focused on intraoperative precision through robotics, navigation, and implant design, while preoperative decision-making remains comparatively underdeveloped. Accurate estimation of patient-specific risk is central to surgical indications, yet existing tools provide limited resolution for consequential outcomes such as 1-year mortality.

A machine learning model was developed using the TriNetX Research Network. Patients undergoing total knee (TKA) or hip (THA) arthroplasty with verifiable 1-year follow-up were included in the analysis. An XGBoost model incorporating 43 preoperative variables was trained using nested cross-validation with isotonic calibration. Model performance was assessed using discrimination, calibration, and decision curve analysis, and the results were used to derive thresholds for a web calculator based on machine-learning-derived inputs. Feature contributions were assessed using SHapley Additive exPlanations (SHAP).

The observed 1-year mortality rate was 0.574% (n = 1,344). On internal validation, the model demonstrated an Area Under the Receiver Operating Characteristic (AUROC) of 0.761 (fold-level range: 0.744-0.787, standard deviation [SD]: 0.016), a Brier score of 0.006, and an average precision of 0.037. Calibration remained stable across risk deciles, with mean absolute error below 0.15% for D1-D9 and modest overprediction confined to D10 (+ 0.16%). Mortality increased monotonically across stratified risk groups: the highest quintile had a 1.488% mortality rate (2.6-fold above baseline), the highest decile had 2.50% (4.4-fold), and the top 5% had 3.57% (6.2-fold). Feature importance analysis identified age as the dominant predictor (mean |SHAP| = 0.407), followed by male sex (0.218), serum creatinine (0.162), hemoglobin (0.153), and BMI (0.145). Threshold-based risk stratification yielded two clinically relevant operating points: a sensitivity-anchored threshold of 0.235% captured 90.1% of all deaths, while a Youden-optimal threshold of 0.767% achieved 83.9% specificity.

One-year mortality following primary THA and TKA can be estimated with discriminative accuracy using routinely available preoperative variables, with predictive importance attributable primarily to age and markers of physiologic reserve. Supported by these findings and future external validations, machine learning-derived mortality models are amenable to prospective clinical deployment via web-based interfaces, enabling individualized preoperative risk quantification at the point of care.