Machine Learning-Based Prediction of Early Patient-Controlled Analgesia Discontinuation After Total Knee...

Background: Patient-controlled analgesia (PCA) is widely used for postoperative pain control after total knee arthroplasty (TKA). Although postoperative nausea and vomiting (PONV) has been extensively studied, early discontinuation of PCA-representing failure to sustain an opioid-based analgesic strategy-has not been adequately investigated as a distinct, decision-relevant outcome. Methods: We conducted a single-center retrospective observational study of 1188 patients undergoing primary TKA who received PCA for postoperative pain management. The primary aim was to develop interpretable machine learning models for predicting early PCA discontinuation using routinely available perioperative variables. A secondary aim was to evaluate the incremental predictive value of hierarchical feature sets reflecting progressively available clinical information. Results: Early PCA discontinuation occurred in approximately 10% of patients, reflecting a relatively low-frequency clinical event associated with class imbalance. Female sex and PONV-related susceptibility factors, including prior nausea/vomiting intolerance, were more common among patients with early PCA discontinuation. The random forest Step 3 model demonstrated acceptable discriminative performance (AUC: 0.77; PR-AUC: 0.38), good calibration accuracy (Brier score: 0.065), and favorable clinical utility on decision curve analysis. Explainable analyses showed that patient-level susceptibility factors-such as prior intolerance to nausea or vomiting and baseline clinical characteristics-contributed more strongly to early PCA discontinuation than perioperative management or procedural variables. Conclusions: Early PCA discontinuation after TKA represents a distinct, decision-based clinical endpoint that is not captured by symptom-focused outcomes such as PONV. Interpretable machine learning models may help identify patients at increased risk of early PCA discontinuation after TKA, which may support informed counseling and proactive planning of individualized postoperative pain management strategies. However, external validation is required before routine clinical implementation.