Related Experiment Video
Updated: Aug 17, 2025
In Vivo Quantification of Hip Arthrokinematics during Dynamic Weight-bearing Activities using Dual Fluoroscopy
Published on: July 2, 2021
Patient-specific Hip Arthroplasty Dislocation Risk Calculator: An Explainable Multimodal Machine Learning-based
Bardia Khosravi1, Pouria Rouzrokh1, Hilal Maradit Kremers1
1Orthopedic Surgery Artificial Intelligence Laboratory, Department of Orthopedic Surgery (B.K., P.R., H.M.K., D.R.L., Q.J.J., M.J.T., C.C.W.), Radiology Informatics Laboratory, Department of Radiology (B.K., P.R., S.F., B.J.E.), Department of Quantitative Health Sciences (H.M.K., D.R.L., W.K.K.), Alix School of Medicine (Q.J.J.), Department of Orthopedic Surgery (R.J.S., M.J.T., C.C.W.), and Department of Clinical Anatomy (C.C.W.), Mayo Clinic, 200 First St SW, Rochester, MN 55905.
This study developed a machine learning model using hip X-rays and patient data to predict dislocation risk after total hip arthroplasty (THA). The multimodal model significantly improved risk prediction compared to clinical data alone.
Area of Science:
- Orthopedic Surgery
- Medical Imaging
- Machine Learning
Background:
- Dislocation is a significant complication following primary total hip arthroplasty (THA).
- Accurate prediction of patient-specific dislocation risk is crucial for surgical planning and improving patient outcomes.
Purpose of the Study:
- To develop and validate a multimodal machine learning pipeline for predicting patient-specific 5-year dislocation risk after primary THA.
- To integrate preoperative radiographic imaging with clinical data for enhanced risk stratification.
Main Methods:
- Retrospective analysis of 17,073 primary THA patients, with a held-out test set of 1,718.
- Development of a hybrid EfficientNet-B4 and Swin-B transformer network for feature extraction from pelvic radiographs.
- Training a multimodal survival XGBoost model using extracted imaging features and clinical characteristics (demographics, comorbidities, surgical details).
- Evaluation using the C-index and Shapley additive explanation (SHAP) values for model interpretability.
Main Results:
- The multimodal model achieved a C-index of 0.74 (95% CI: 0.69, 0.78), significantly outperforming a clinical-only model (C-index: 0.64; 95% CI: 0.60, 0.68; P = .02).
- The 5-year dislocation incidence in the study cohort was 2%.
Conclusions:
- The developed multimodal machine learning model demonstrates superior discrimination ability in predicting THA dislocation risk.
- This explainable risk calculator can serve as a valuable tool for preoperative risk stratification and surgical planning in THA.
- Integration of imaging features with clinical data enhances the predictive performance for THA complications.
