Development and internal validation of machine learning models with post hoc explainability analyses for predicting functional outcomes following total knee arthroplasty: a retrospective cohort study
Year: 2026
Type: article
Abstract: To develop and internally validate a machine learning (ML) framework with post hoc explainability analyses for exploring the predictability of early functional outcomes at 3 months following total knee arthroplasty (TKA), and to characterize model- and case-specific feature contributions for two clinically selected, objectively measured outcomes representing walking endurance and knee extensor strength. This retrospective cohort study included 321 patients for the 6-minute walk test (6MWT) prediction model and 316 patients for the knee extension peak torque normalized to body weight (Knee Ex PT(BW)) prediction model using complete-case analysis after exclusion of cases with missing data. Nine postoperative functional and patient-reported outcomes were assessed at baseline and at 3 months after surgery, including pain, mobility, quality of life, walking endurance, and knee muscle strength. Pre–post comparisons were performed to characterize overall recovery patterns. Exploratory screening of nine candidate postoperative outcomes was performed using seven ML algorithms with nested cross-validation. Detailed post hoc explainability analyses were subsequently focused on the 6MWT and Knee Ex PT(BW), which were retained on clinical rather than statistical grounds because they represent complementary, objectively measured domains of postoperative recovery; this outcome selection was exploratory and not prespecified. Spearman correlation analysis was applied to examine structural relationships among selected outcomes. Model diagnostics and permutation-based feature importance analyses were subsequently performed. ML model performance was evaluated using mean absolute error, root mean squared error, and coefficient of determination (R²), with error metrics additionally reported in original clinical units. To assess the incremental predictive value of ML, ordinary least squares (OLS), ridge, lasso, and elastic-net regression models were evaluated as conventional comparators under the same analytical framework, with paired bootstrap estimation of differences in R². Local Interpretable Model-agnostic Explanations (LIME) were applied to characterize case-specific model behavior. The mean age was 71.9 ± 5.9 years in the 6MWT cohort and 71.8 ± 5.9 years in the Knee Ex PT(BW) cohort, with male patients accounting for 15.3% ( n = 49) and 15.5% ( n = 49) of each cohort, respectively. Both the 6MWT and Knee Ex PT(BW) demonstrated significant postoperative improvement at 3 months (all p < 0.001). Exploratory screening showed that several outcomes not retained for detailed analysis achieved explanatory power comparable to or higher than that of the two retained outcomes. The 6MWT and Knee Ex PT(BW) were therefore retained on clinical rather than statistical grounds. For the 6MWT, the TabPFN Regressor achieved an R² of 0.283 and an RMSE of 73.614 m, whereas the Support Vector Machine Regressor achieved an R² of 0.283 and an RMSE of 21.644%BW for Knee Ex PT(BW). Overall predictive performance was modest. Paired bootstrap comparisons with the best-performing conventional regression models showed no clear incremental predictive advantage of ML. In the unified analysis, the highest permutation-importance values for the 6MWT model were observed for preoperative WOMAC Total, age, and preoperative 6MWT distance, whereas sex and preoperative knee extensor peak torque had the highest values in the Knee Ex PT(BW) model. Permutation-importance and LIME analyses yielded model- and case-specific feature contributions, respectively, without implying biological or causal importance. Machine learning with post hoc explainability analyses provided exploratory insights into the predictability of early functional outcomes following TKA. Among nine candidate outcomes, the 6MWT and Knee Ex PT(BW) were retained for detailed analysis on clinically motivated, post hoc grounds rather than because they were the best-performing prediction targets. Overall predictive performance was modest, and the current models are insufficient for individual-level clinical decision-making. No clear incremental predictive advantage of ML over conventional regression was demonstrated; this does not establish equivalence between the modeling approaches. The findings may have been substantially influenced by missing-data-related selection effects because complete-case analysis retained only approximately one quarter of the screened population. They should therefore be regarded as exploratory and hypothesis-generating, particularly given the single-center design and absence of external validation. Validation in larger, independent multicenter cohorts is required before clinical implementation can be considered.
Language: English
Subfield: Surgery
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $3,090
OpenAlex ID: W7222967556
FWCI:
Cites:
Cited by:
Related to: