Behavioral and Engagement Predictors of Arthroplasty Clinic No-Shows: A Calibrated Machine Learning Analysis

Background Missed clinic visits in arthroplasty care can impair preoperative optimization, delay surgery, and disrupt follow-up. Drivers of arthroplasty-specific no-shows remain incompletely characterized. Methods We extracted 94,723 scheduled arthroplasty appointments at a single tertiary academic center (March 2013–May 2025), retaining each patient's most recent encounter (N = 17,614) for primary analyses. Demographic, socioeconomic, clinical, scheduling, and prior-attendance variables were drawn from the electronic health record. Multivariable logistic regression and isotonic-calibrated machine learning models estimated no-show risk. Discrimination, calibration, and classification were assessed at multiple operating thresholds, and feature contributions were examined via Shapley Additive exPlanations. Results The no-show rate was 6.9%. Prior attendance behavior was the strongest predictor: each 10-percentage-point increase in historical no-show proportion conferred ∼26% higher odds (adjusted odds ratio [aOR]: 1.26 per 10 pp; full-range aOR: 9.86; 95% confidence interval: 6.16–15.76; P < .001), and appointment confirmation was strongly protective (aOR: 0.24; 95% confidence interval: 0.19–0.29; P < .001). Arthroplasty-relevant diagnoses such as osteoarthritis were associated with lower odds, while social determinants of health showed modest but persistent associations with increased risk. Calibrated machine-learning models achieved good discrimination (area under the receiver operating characteristic curve: 0.80) and calibration (Brier skill +0.12); at a capacity-matched threshold flagging the highest-risk 20%, the best-performing model captured ∼60% of no-shows with 82% specificity. Shapley Additive exPlanations highlighted prior no-show history and confirmation as dominant features. Conclusions Arthroplasty clinic no-shows reflects engagement patterns, administrative processes, and patient context. Calibrated, interpretable machine-learning models capturing ∼60% of no-shows within the highest-risk 20% of encounters support prospective evaluation for targeted, supportive outreach with equity safeguards.