Scholarlink

Background <p>Fixation choice (polymethyl-methacrylate bone cement or cementless implants) in total knee arthroplasty (TKA) varies widely across surgeons, reflecting differences in clinical judgment. While traditional statistical approaches are commonly used to study surgical decision-making, they may be limited in capturing complex and multifactorial relationships inherent in such data. This study evaluates how different supervised machine learning (SML) algorithms characterize fixation choice and prioritize predictors, rather than developing a clinical decision-support tool.</p> Methods <p>We analyzed data from the multicenter Patient-Centered Outcomes Research Institute (PCORI)-funded Comparative Effectiveness of Pulmonary Embolism Prevention after Hip and Knee Replacement (PEPPER) trial, linked with the American Hospital Association Annual Survey. Adult patients undergoing elective primary TKA for osteoarthritis between December 2016 and May 2024 were included. The primary outcome was the binary classification of cemented versus cementless fixation. Five models were evaluated (logistic regression, LASSO, support vector machine, XGBoost, and Random Forest). Model performance was assessed using discrimination, precision–recall performance, calibration, and F1 score within a nested cross-validation framework. All confidence intervals were estimated using surgeon-clustered bootstrap resampling. Permutation importance was used for cross-model feature comparison, and SHapley Additive exPlanations (SHAP) were applied to interpret the best-performing model.</p> Results <p>Among 7848 patients treated by 140 surgeons at 29 hospitals, 85.8% underwent cemented TKA. RF achieved the highest discriminatory performance (AUROC = 0.868; 95%CI:0.826–0.891, F1 = 0.944; 95%CI:0.927–0.951) and the lowest Brier score (0.080; 95%CI: 0.072–0.091). In the RF model, patient BMI and age emerged as the most influential predictors of fixation choice, with SHAP analyses indicating an inverse association between BMI and cemented fixation and a non-linear relationship between age and the likelihood of cemented TKA. In contrast, non-tree-based models more frequently prioritized hospital-level and geographic characteristics over patient factors.</p> Conclusions <p>Tree-based SML models, particularly RF, showed the highest performance in this cohort and prioritized patient-level predictors more consistently than non–tree-based models. Substantively, fixation choice within this study cohort was more strongly associated with patient-specific factors such as BMI and age. These findings may help inform understanding of how patient characteristics are associated with fixation choice in current clinical practice.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative analysis of supervised machine learning models for predicting surgeon fixation choice in total knee arthroplasty

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月11日

查看全文
  • Hyunkyu Ko,
  • Brook I. Martin,
  • Christopher E. Pelt,
  • Vincent D. Pellegrini Jr.

摘要

Background

Fixation choice (polymethyl-methacrylate bone cement or cementless implants) in total knee arthroplasty (TKA) varies widely across surgeons, reflecting differences in clinical judgment. While traditional statistical approaches are commonly used to study surgical decision-making, they may be limited in capturing complex and multifactorial relationships inherent in such data. This study evaluates how different supervised machine learning (SML) algorithms characterize fixation choice and prioritize predictors, rather than developing a clinical decision-support tool.

Methods

We analyzed data from the multicenter Patient-Centered Outcomes Research Institute (PCORI)-funded Comparative Effectiveness of Pulmonary Embolism Prevention after Hip and Knee Replacement (PEPPER) trial, linked with the American Hospital Association Annual Survey. Adult patients undergoing elective primary TKA for osteoarthritis between December 2016 and May 2024 were included. The primary outcome was the binary classification of cemented versus cementless fixation. Five models were evaluated (logistic regression, LASSO, support vector machine, XGBoost, and Random Forest). Model performance was assessed using discrimination, precision–recall performance, calibration, and F1 score within a nested cross-validation framework. All confidence intervals were estimated using surgeon-clustered bootstrap resampling. Permutation importance was used for cross-model feature comparison, and SHapley Additive exPlanations (SHAP) were applied to interpret the best-performing model.

Results

Among 7848 patients treated by 140 surgeons at 29 hospitals, 85.8% underwent cemented TKA. RF achieved the highest discriminatory performance (AUROC = 0.868; 95%CI:0.826–0.891, F1 = 0.944; 95%CI:0.927–0.951) and the lowest Brier score (0.080; 95%CI: 0.072–0.091). In the RF model, patient BMI and age emerged as the most influential predictors of fixation choice, with SHAP analyses indicating an inverse association between BMI and cemented fixation and a non-linear relationship between age and the likelihood of cemented TKA. In contrast, non-tree-based models more frequently prioritized hospital-level and geographic characteristics over patient factors.

Conclusions

Tree-based SML models, particularly RF, showed the highest performance in this cohort and prioritized patient-level predictors more consistently than non–tree-based models. Substantively, fixation choice within this study cohort was more strongly associated with patient-specific factors such as BMI and age. These findings may help inform understanding of how patient characteristics are associated with fixation choice in current clinical practice.