Abstract / Summary
Abstract Introduction Same-day and early discharge after total hip and knee arthroplasty (THA/TKA) depend not only on patient characteristics but on the perioperative pathway itself. This study used supervised machine learning (ML) to evaluate a universal, non-selective enhanced recovery after surgery (ERAS) pathway at a high-volume tertiary centre, identifying the perioperative factors most strongly associated with failure of Day 0 or Day 1 discharge, and of same-day (Day 0) discharge. Methods All patients undergoing primary THA, TKA, or partial knee replacement (PKR) within a universal, non-selective ERAS pathway (April 2024–May 2025) were included, with no exclusion on the basis of age, comorbidity, or anticipated day-case suitability. ML models (Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine, Decision Tree) were developed in Python using the full set of preoperative, anaesthetic, and postoperative pathway variables, with SHAP used for interpretation. Results A total of 3,025 patients were included (median age 70.1 years, IQR 62.3–76.6). With discharge failure as the positive class, test ROC-AUC for failure of Day 0 or Day 1 discharge ranged from 0.66 to 0.76 (PR-AUC 0.34–0.50; baseline failure prevalence 18%). For same-day discharge failure, ROC-AUC ranged from 0.76 to 0.82 (F1 0.76–0.83; PR-AUC 0.89–0.93; baseline prevalence 73%). SHAP analysis ranked the factors most strongly associated with discharge failure. For same-day discharge, time to mobilisation was the most significant factor, followed by postoperative radiography before return to the ward and mobiliser designation. For Day 0 or Day 1 discharge, the strongest associations were ASA grade, sex, operative side, and time to mobilisation. Conclusion Supervised ML is an effective framework for evaluating a non-selective ERAS pathway. Time to mobilisation dominated same-day discharge and was a leading factor for Day 0 or Day 1 discharge; it warrants prospective confirmation as a modifiable target.