JOA - 2026-07-14 - Journal Article
Predicting Patient-Reported Outcome Measures, Satisfaction, Healthcare Utilization, Mortality, and Return to Work After Total Knee Arthroplasty Using Machine Learning: A 14,900-Patient Study.
Khan ST, Sharma A, Patel SV, Jevnikar BE, Rotroff D, Cleveland Clinic Center of Excellence in Osteoarthritis Group, Piuzzi NS
Topics
Key Takeaway
Random Forest and XGBoost models predicted 1-year mortality with 73% accuracy and return-to-work with 78% accuracy in 14,900 primary TKA patients, with baseline KOOS Joint Replacement score, Area Deprivation Index, and race among the top predictors across outcomes.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study developed and validated Random Forest and XGBoost machine learning models using demographic, clinical, socioeconomic, and surgical variables from 14,900 primary unilateral TKA patients at a single tertiary center (2016–2022) to predict KOOS subscales, Patient Acceptable Symptom State, length of stay, discharge disposition, 90-day readmission, 1-year mortality, and return to work. KOOS outcome prediction yielded RMSE of 15.13–23.49; categorical outcome accuracy ranged from 55–71% for healthcare utilization metrics, 73% for mortality, and 78% for return to work. Area Deprivation Index and race emerged as significant predictors alongside established clinical variables, highlighting the independent contribution of socioeconomic factors.
Key Limitation
Single-center derivation without external validation means model performance metrics likely overestimate real-world predictive accuracy across diverse health systems and patient populations.
Original Abstract
BACKGROUND
Patient-reported outcome measures and healthcare utilization metrics are increasingly used to evaluate the success of total knee arthroplasty (TKA). Predictive models may improve preoperative planning, patient counseling, and resource allocation. This study developed and validated machine learning models to predict postoperative patient-reported outcomes, satisfaction, healthcare utilization, mortality, and return to work after primary TKA.
METHODS
A prospective cohort of 14,900 patients who underwent primary unilateral TKA at a large tertiary academic center from 2016 to 2022 was analyzed. Patients undergoing bilateral TKA or missing baseline patient-reported outcome measures were excluded. Random Forest and XGBoost models were developed using baseline demographic, clinical, socioeconomic, and surgical variables. Outcomes included postoperative Knee injury and Osteoarthritis Outcome Score (KOOS) Pain, Physical Function, Joint Replacement, and Quality of Life subscales; Patient Acceptable Symptom State; length of stay; discharge disposition; 90-day readmission; 1-year mortality; and return to work. Model performance was evaluated using root mean square error for continuous outcomes and accuracy for categorical outcomes.
RESULTS
Predictive performance was moderate to strong across outcomes. For KOOS outcomes, root mean square error ranged from 15.13 to 23.49. Accuracy for healthcare utilization outcomes ranged from 55 to 71%. Accuracy reached 73% for 1-year mortality and 78% for return to work. Important predictors across models included baseline KOOS Joint Replacement score, patient-reported outcome phenotype, age, body mass index, Area Deprivation Index, race, and surgery start time. The Charlson Comorbidity Index also contributed to the prediction of select clinical outcomes.
CONCLUSION
Machine learning models demonstrated moderate to strong performance for predicting patient-reported outcomes, satisfaction, healthcare utilization, mortality, and return to work after TKA. Integration of these tools may allow population-level variable parsing - identifying distinct socioeconomic, functional, and operational signals within an optimized cohort. External validation is needed before widespread clinical implementation.