A Radiomics-Based Machine Learning Model to Predict Total Knee Arthroplasty Failure from Plain Radiographs

Area of Science:

• Orthopedic Surgery
• Medical Imaging
• Artificial Intelligence

Background:

• Total knee arthroplasty (TKA) revisions are increasing, with poorer outcomes than primary procedures.
• Early detection of primary TKA failure is critical for improving patient outcomes and managing healthcare resources.
• Current methods for TKA failure detection can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a radiomics-based machine learning model for automated detection of TKA failure.
• To assess the model's performance using standard metrics like F1-score, balanced accuracy, and AUC.
• To provide a tool that assists clinicians in identifying TKA failure early from standard radiographs.

Main Methods:

• A dataset of 95 TKA patient radiographs (44 failed, 51 non-failed) was utilized.
• 465 radiomic features were extracted after preprocessing.
• A 100-fold cross-validation procedure was implemented, evaluating feature selection (LASSO) and classifiers (Logistic Regression).

Main Results:

• The LASSO feature selection combined with Logistic Regression achieved the best performance.
• The model demonstrated an F1-score of 0.701, balanced accuracy of 0.710, and AUC of 0.783.
• These results indicate the model's potential for accurate TKA failure detection.

Conclusions:

• The developed radiomics-based machine learning model shows significant potential for automatically detecting TKA failure from plain radiographs.
• This automated approach can aid clinicians by reducing workload and minimizing inter- and intra-observer variability.
• Early and accurate detection of TKA failure can lead to timely interventions and improved patient management.