XGBoost model for TKA complications: moderate discrimination for major complications (AUC 0.68), no better than chance for residual pain (AUC 0.53)
Source: Journal of Orthopaedics·Published: 2026
Authors: Müller D, Gillani A, Hinterwimmer F, Arber A, Graichen H, von Eisenhart-Rothe R, Lazic I·DOI: 10.1016/j.jor.2025.08.038Open Access
Key figure: Figure 3 — ROC curves for the three target variables, showing the AUC of 0.68 for major complications, 0.65 for any complication, and 0.53 for residual pain. View in source
Bottom line: Trained on 783 primary TKAs, the XGBoost model achieved AUC 0.68 for major complications requiring revision, AUC 0.65 for any complication, and AUC 0.53 for one-year residual pain (Visual Analog Scale ≥ 4). Smoking status, ASA ≥ 3, and previous open reduction internal fixation of the knee emerged as the dominant features for complication prediction.
What the study did
The authors retrospectively analyzed 783 patients who underwent primary total knee arthroplasty at Klinikum Rechts der Isar, a single academic center affiliated with Technical University of Munich. Demographic, surgical, and outcome data covering 12 AAHKS-defined preoperative risk factors were extracted from the institutional database. An XGBoost gradient-boosted tree model was trained to perform binary classification for three target variables: major complications requiring revision, any complication (major or minor), and residual pain at one-year follow-up, defined as Visual Analog Scale score 4 or higher. SMOTE oversampling addressed class imbalance. Model performance was evaluated with area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. SHAP values were computed to rank feature importance.
What they found
The model discriminated major complications requiring revision with an AUC of 0.68 and any complication with an AUC of 0.65, both categorized by the authors as moderate. Residual pain at one year was not reliably predicted, with an AUC of 0.53 (at or near chance). Among the 12 AAHKS-defined preoperative risk factors, the most important predictors identified by SHAP analysis were smoking status, ASA physical status class of 3 or higher, and a history of previous open reduction and internal fixation of the knee. The 83.8% one-year follow-up rate was above typical registry standards.
Why it matters for orthopedic practice
Preoperative risk prediction is increasingly embedded in shared decision-making conversations and in bundled-payment models that adjust for patient complexity. A moderate AUC of 0.68 for major complications is aligned with prior published models and suggests that complications research has plateaued at the level of risk stratification useful for counseling rather than precise individual prediction. The flat performance on residual pain is an important result in its own right: structured preoperative risk factors alone cannot predict one-year pain outcomes, and the field will need richer inputs (imaging features, patient-reported baselines, biomarkers) if meaningful pain prediction is a clinical goal.
Limitations
This is a retrospective single-center study at an academic referral hospital, which limits generalizability to community practice and to other healthcare systems. Class imbalance was significant, with major complications occurring in fewer than 3% of patients, and the authors used SMOTE oversampling to mitigate this, a technique that can inflate reported model performance relative to deployment performance. External validation on a population from a different institution was not performed. The follow-up window for residual pain was one year, and longer-term pain trajectories were not modeled. The model was trained on AAHKS-defined structured variables only, so the ceiling reported here may not reflect what a richer feature set could achieve.
Müller D, Gillani A, Hinterwimmer F, Arber A, Graichen H, von Eisenhart-Rothe R, Lazic I. Machine learning-based prediction of complications and residual pain after total knee arthroplasty. J Orthop. 2026;71:60-66. doi:10.1016/j.jor.2025.08.038
Publishing AI research in orthopedics?
OSCRSJ accepts case reports and series on novel AI-assisted diagnoses and surgical planning. Free to publish in 2026.
Submit a manuscriptOSCRSJ News items are editorial summaries for educational purposes. They are not clinical recommendations, endorsements, or substitutes for the primary literature. Always consult the source paper and applicable specialty-society guidelines before changing practice.