Interpretable machine learning model for predicting postoperative hypoproteinemia in middle-aged and elderly patients following total joint arthroplasty: a multicenter retrospective study.

Background: Postoperative hypoproteinemia is a prevalent yet frequently overlooked complication following total hip or knee arthroplasty (TJA), adversely impacting wound healing and postoperative recovery in patients. This study aims to identify and evaluate the risk factors associated with hypoproteinemia in middle-aged and elderly patients after TJA, as well as to develop and validate an interpretable machine learning model for predicting high-risk patients promptly. Materials and method: This study employed a multicenter retrospective research design, gathering clinical data from middle-aged and elderly patients who underwent TJA surgery across two hospitals. Key predictors were identified and selected using minimum absolute shrinkage and selection operator (LASSO) regression. The predictive performance of eight machine learning algorithms was compared, and the SHapley Additive exPlanations (SHAP) analysis was conducted to rank the features and assess the interpretability of the optimal model. Results: A total of 1,383 participants were evaluated, including 585 patients (mean age: 65 ± 10; 233 men) in the training cohort, 250 patients (mean age: 65 ± 9; 101 men) in the internal validation cohort, and 548 patients (mean age: 69 ± 10; 186 men) in the external validation cohort. Key predictors identified include age, surgery time, anesthesia method, erythrocyte sedimentation rate (ESR), and serum calcium level. Notably, the Light Gradient Boost Machine (LightGBM) model exhibited the highest predictive performance, with AUC values of 1.000, 0.968, and 0.855 in the training, internal validation, and external validation sets, respectively. Consequently, an interpretable LightGBM model incorporating five variables was established. SHAP analysis showed that age, ESR, surgery time, serum calcium level and anesthesia method made significant contributions to the prediction of postoperative hypoproteinemia. Finally, we present the model as an accessible web-based tool for individualized, real-time clinical decision-making (https://hypoproteinemia-prediction-uzhpp3jadq8ryey4jxq3qe.streamlit.app/). Conclusion: This study successfully developed and validated a machine learning prediction model that exhibit high accuracy and interpretability. This model demonstrates high accuracy and interpretability in predicting the risk of hypoproteinemia following TJA and elucidates intricate nonlinear relationships among various factors. It may serve as a promising tool for early risk screening and proactive prevention, though further prospective validation is needed before routine clinical implementation.