Development of Novel and Interpretable Automated Models for Predicting Total Knee Replacement in Knee Osteoarthritis: An International Multicenter Study.
Authors
Affiliations (9)
Affiliations (9)
- Center for Rehabilitation Medicine, Department of Radiology, Zhejiang Provincial People's Hospital (Affiliated People's Hospital), Hangzhou Medical College, Hangzhou, China (F.Z., L.X., C.L., D.H.).
- Department of Radiology, Affiliated XiHu Hospital Hangzhou Medical College, Hangzhou, China (Y.C.).
- Nanfang Hospital, Southern Medical University, Guangzhou, China (X.Z.).
- Department of Radiology, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China (X.C.).
- The First School of Clinical Medicine, Zhejiang Chinese Medical University, Hangzhou, China (C.H., F.Y.).
- School of Medicine, South China University of Technology, Guangzhou, China (Y.S., S.H.).
- KnowX Tech Inc. Toronto, ON, Canada (R.G.).
- Department of Radiology, Jiaxing Hospital of Traditional Chinese Medical, Jiaxing, China (M.H.).
- Center for Rehabilitation Medicine, Department of Radiology, Zhejiang Provincial People's Hospital (Affiliated People's Hospital), Hangzhou Medical College, Hangzhou, China (F.Z., L.X., C.L., D.H.). Electronic address: [email protected].
Abstract
Subtle changes in the infrapatellar fat pad (IPFP) enable prediction and quantification of knee osteoarthritis (KOA). This study aimed to develop a bimodal model for predicting the risk of KOA progression to total knee replacement (TKR). A total of 4039 eligible participants from the Osteoarthritis Initiative (OAI) database were retrospectively enrolled in this study, who were further divided into the training, testing, and internal validation cohorts. Additionally, datasets from the Center D and Center E hospitals were collected as external validation cohorts I and II. Based on the optimal radiomic features of the IPFP and KL grading factors, the Fatpad-Score, KL model, and a bimodal model integrating both were developed, respectively. The predictive performance of models was evaluated using the area under the receiver operating characteristic curve (ROC-AUC) and diagnostic metrics. Compared with the Fatpad-score (AUC = 0.848; 95% confidence interval [95% CI]: 0.773-0.916) and KL grading model (AUC = 0.801; 95% CI: 0.726-0.868), the bimodal model outperformed both in internal validation, achieving an AUC of 0.903 (95% CI: 0.846-0.950) and accuracy of 0.871. Pairwise comparison of AUC via the net reclassification improvement (NRI) test further confirmed that the bimodal model had enhanced performance (all P < 0.001). The bimodal model achieved accuracies of 0.918 and 1.000 in external validation cohorts I and II, respectively. The proposed bimodal model better predicted the risk of TKR in KOA patients than did the Fatpad-score and the K-L grading model.