Machine Learning Model for Selection of Cementless Total Knee Arthroplasty Candidates Utilizing Patient and Radiographic Parameters.
Authors
Affiliations (5)
Affiliations (5)
- School of Medicine, University of Louisville, Louisville, Kentucky, USA.
- Department of Orthopedic Surgery, University of Louisville, Louisville, Kentucky, USA.
- Department of Radiology, University of Louisville, Louisville, Kentucky, USA.
- UofL Health, Louisville, Kentucky, USA.
- Department of Mechanical Engineering, University of Louisville, Louisville, Kentucky, USA.
Abstract
The use of cementless total knee arthroplasty (TKA) has significantly increased over the past decade. However, there is no objective criteria or consensus on parameters for patient selection for cementless TKA. The purpose of this study was to develop a machine learning model based on patient and radiographic parameters that could identify patients indicated for cementless TKA. We developed an explainable recommendation model using multiple patient and radiographic parameters (BMI, Age, Gender, Hounsfield Units [HU] from CT for density of tibia). The predictive model was trained on medical, operative, and radiographic data of 217 patients who underwent primary TKA. HU density measurements of four quadrants of the proximal tibia were obtained at region of interest on preoperative CT scans. which were then incorporated into the model as a surrogate for bone mineral density. The model employs Local Interpretable Model-agnostic Explanations in combination with bagging ensemble techniques for artificial neural networks. Model testing on the 217-patient cohort included 22 cemented and 38 cementless TKA cases. The model successfully identified 19 cemented patients (sensitivity: 86.4%) and 37 cementless patients (specificity: 97.4%) with an AUC = 0.94. Use of cementless TKA has grown significantly. There are currently no standard radiographic criteria for patient selection. Our machine learning model demonstrated 97.4% specificity and should improve with more training data. Future improvements will include incorporating additional cases and developing automated HU extraction techniques.