Related Experiment Videos
Development of Machine Learning Algorithms Predicting Psychological Distress After Total Joint Arthroplasty
Michelle M Ramirez1, Maggie E Horn2, Steven Z George3
1Duke University School of Medicine, Department of Population Health Sciences, Durham, NC, USA; Duke University School of Medicine, Department of Orthopaedic Surgery, Durham, NC, USA; University of North Carolina-Chapel Hill, Thurston Arthritis Research Center, Department of Medicine, Chapel Hill, NC, USA.
Background:
Psychological distress is associated with suboptimal outcomes after total joint arthroplasty (TJA). This study aimed to develop and evaluate machine learning (ML) models to predict a high psychological distress phenotype using only preoperative data.
Methods:
We conducted a retrospective secondary analysis of patients undergoing primary hip or knee arthroplasty at Duke University between 2018 and 2024. Phenotypes were derived using latent class analysis of the Optimal Screening for Prediction of Referral and Outcome Yellow Flag (OSPRO-YF) tool. There were four models (1) elastic net, (2) XGBoost, (3) random forest, and (4) logistic regression trained to predict a high-distress phenotype using preoperative demographic, clinical, and patient-reported data. The dataset was split into training and testing sets (70:30). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, Brier score, sensitivity, specificity, calibration, and decision curve analysis.
Results:
A total of 494 patients (64% women) undergoing TJA (knee 57%, hip 43%) were included; 18% (n = 89) were classified as having high postoperative distress. Patients in the high-distress phenotype demonstrated lower Patient Reported Outcomes Measurement Information System (PROMIS) physical function, higher PROMIS pain interference, pain ratings, and a greater prevalence of high-impact chronic pain. The elastic net model performed best, with an AUC of 0.75 (95% CI [confidence interval]: 0.62 to 0.85), compared with logistic regression (0.73), random forest (0.71), and XGBoost (0.65). Key preoperative predictors of high distress included a higher preoperative OSPRO-YF count, greater pain, preoperative depression, and higher body mass index.
Conclusion:
Commonly collected data from routine preoperative clinical care show promise for predicting psychological distress after TJA. The elastic net model demonstrated the strongest overall performance and interpretability. Such models could support early identification of patients at risk for postoperative psychological distress, enabling targeted behavioral health referral or psychologically informed physical therapy prior to surgery. Future work should validate these models in larger, multi-site cohorts.