Afshin Jamshidi1,2, Jean-Pierre Pelletier1, Aurelie Labbe3, François Abram4, Johanne Martel-Pelletier1, Arnaud Droit2. 1. Osteoarthritis Research Unit, University of Montreal Hospital Research Centre (CRCHUM), Montreal, Quebec, Canada. 2. Laval University Hospital Research Centre, Quebec, Canada. 3. Department of Decision Sciences, HEC Montreal, Montreal, Quebec, Canada. 4. Medical Imaging Research and Development, ArthroLab Inc, Montreal, Quebec, Canada.
Abstract
OBJECTIVE: By using machine learning (ML), our study aimed to build a model to predict risk and time to total knee replacement (TKR) of an osteoarthritic knee. METHODS: Features were from OAI at baseline. Lasso's Cox identified the ten most important among 1107 features. The prognostic power of the selected features was assessed by Kaplan-Meier and applied to seven ML methods: Cox, DeepSurv, Random Forest, linear/kernel Support Vector Machine (SVM), and linear/neural Multi-Task Logistic Regression. As some of the ten features included similar X-ray measurements, we further looked at using the least features without compromising the accuracy of the model. Prediction performance was assessed by the concordance index (C-index), Brier score, and time-dependent area under the curve (AUC). RESULTS: Identified features included X-rays, the MRI feature bone marrow lesions (BML) in medial condyle, hyaluronic acid injection, performance measure, medical history, and knee symptoms. The methodologies Cox, DeepSurv, and linear SVM demonstrated the highest accuracy (C-index, 0.85, Brier score 0.02, and AUC, 0.87). DeepSurv was chosen to build the prediction model to estimate the time to TKR for a given knee. Moreover, we were able to use only three features (C-index, 0.85, Brier score 0.02, and AUC, 0.86) including BML, KL grade and knee symptoms, to predict risk and time of a TKR event. CONCLUSION: For the first time, using the OAI cohort, we developed a model to predict with high accuracy if and when a given osteoarthritic knee would require TKR and who would likely progress fast toward this event. This article is protected by copyright. All rights reserved.
OBJECTIVE: By using machine learning (ML), our study aimed to build a model to predict risk and time to total knee replacement (TKR) of an osteoarthritic knee. METHODS: Features were from OAI at baseline. Lasso's Cox identified the ten most important among 1107 features. The prognostic power of the selected features was assessed by Kaplan-Meier and applied to seven ML methods: Cox, DeepSurv, Random Forest, linear/kernel Support Vector Machine (SVM), and linear/neural Multi-Task Logistic Regression. As some of the ten features included similar X-ray measurements, we further looked at using the least features without compromising the accuracy of the model. Prediction performance was assessed by the concordance index (C-index), Brier score, and time-dependent area under the curve (AUC). RESULTS: Identified features included X-rays, the MRI feature bone marrow lesions (BML) in medial condyle, hyaluronic acid injection, performance measure, medical history, and knee symptoms. The methodologies Cox, DeepSurv, and linear SVM demonstrated the highest accuracy (C-index, 0.85, Brier score 0.02, and AUC, 0.87). DeepSurv was chosen to build the prediction model to estimate the time to TKR for a given knee. Moreover, we were able to use only three features (C-index, 0.85, Brier score 0.02, and AUC, 0.86) including BML, KL grade and knee symptoms, to predict risk and time of a TKR event. CONCLUSION: For the first time, using the OAI cohort, we developed a model to predict with high accuracy if and when a given osteoarthritic knee would require TKR and who would likely progress fast toward this event. This article is protected by copyright. All rights reserved.
Entities:
Keywords:
X-ray; feature selection; machine learning; magnetic resonance imaging; osteoarthritis; prediction; survival analysis; total knee replacement