Development and Validation of the APEX-HBD SCORe: A Multivariable Prediction Model for Postoperative Complications in Patients Undergoing Shoulder Arthroplasty.

Background: The incidence of shoulder arthroplasty (SA) has risen significantly, driven by expanded indications. This study aims to derive and validate a model for classifying patients based on the risk of short-term complications using logistic regression (LR) and other machine learning (ML) techniques.

Methods: We analyzed de-identified data from the American College of Surgeons' NSQIP database (2005-2022), identifying 39,028 patients who underwent SA using CPT codes. The dataset was split into derivation (60%) and validation (40%) cohorts. We constructed baseline classifiers for complications using backward stepwise multivariate LR and developed the APEX-HBD SCORe, a point-based system to stratify patients into low (<5%), moderate (5-11%), and high-risk (≥19%) categories. To improve accuracy, we also developed ML models, including Gradient Boosting, AdaBoost, Random Forest, and Extra Randomized Trees, using the same predictors identified in the LR model.

Results: The derivation cohort (23,417 patients) reported 1,476 (6.3%) patients with complications. LR identified 11 predictors, including albumin levels, hematocrit, ASA classification, preoperative transfusion, and other relevant factors. LR achieved AUCs of 72%, 75%, 77%, and 63% for any, medical, serious medical, and surgical complications, respectively, in the derivation cohort, and 70%, 75%, 73%, and 58% in the validation cohort-outperforming the 5-item modified frailty index (MFI_5). Gradient Boosting performed best among ML models, with AUCs of 73%, 82%, 76%, and 67%. APEX-HBD SCORe risk stratification revealed a progressive increase in complication rates across categories, confirmed in the validation cohort.

Conclusion: This 16-year analysis introduces the APEX-HBD SCORe, a validated ML-augmented tool predicting 30-day complications using 11 patient factors. It aids in patient stratification, counseling, preoperative planning, and tailored postoperative management.

Level Of Evidence: Basic Science Study; Development and Validation of Predictive Model using Machine Learning.