An uncertainty-aware approach to postoperative knee satisfaction prediction after total knee arthroplasty: nested cross-fitted ensemble learning with conformal prediction intervals

Prediction of continuous patient-reported outcomes from limited tabular clinical data is challenging because model rankings may vary across learning algorithms and predictive uncertainty is not captured by point estimates alone. This study evaluated postoperative knee satisfaction using individual regression learners, nested cross-fitted ensemble methods, and distribution-free prediction intervals. An analytic dataset of 98 observations was evaluated using Ridge, Elastic Net, Random Forest, XGBoost, and CatBoost. Four ensemble strategies were considered: uniform averaging, inverse-MAE weighting, Ridge stacking, and a convex Super Learner. Predictive performance was assessed using five-fold outer cross-validation. Within each outer-training set, a separate five-fold inner cross-validation procedure was used to estimate all data-dependent ensemble weights and meta-learner coefficients. Split-conformal prediction was embedded within the outer evaluation to generate finite-sample corrected 95% prediction intervals. SHAP and a perturbation-based local surrogate analysis were used for post-hoc interpretation of full-data base-learner refits. Random Forest achieved the lowest overall MAE (2.150) and MAPE (8.603%), whereas the cross-fitted convex Super Learner yielded the lowest RMSE (3.257) and the highest \(\:{R}^{2}\) (0.792). Its numerical advantage over Ridge was small, with Ridge obtaining an RMSE of 3.280 and an \(\:{R}^{2}\) of 0.789. Among the ensemble methods, inverse-MAE weighting produced the lowest MAE (2.168) and MAPE (8.766%), while uniform averaging showed similar performance. Empirical conformal coverage ranged from 0.949 to 0.980. Ridge generated the narrowest mean prediction intervals overall (17.683), whereas the convex Super Learner generated the narrowest intervals among the ensemble methods (19.812). The SHAP and local surrogate analyses primarily attributed predictions to KSS-related subscale variables. Nested cross-fitting prevented the second-stage ensemble parameters from being estimated and evaluated on the same observations. No single learning method consistently dominated all performance measures, and the observed advantages were small and metric-dependent. The framework provides an internally evaluated approach for combining point prediction with marginal uncertainty quantification, but external validation with larger cohorts and clinically independent predictors is required.