Background/Purpose: Current assessment of knee osteoarthritis (OA) is primarily based on a patient’s personal and familial history, clinical features, and on radiographies. However, such information does not provide enough evidence to lead to a robust prediction/prognosis of OA fast progressors. This study aims to identify early significant predictors of knee OA progression using advanced machine learning (ML) and deep learning (DL) algorithms and propose a prediction model based on a short number of selected predictors and outcomes.

Methods: We used six feature selection models, including LASSO (Least Absolute Shrinkage and Selection Operator), Elastic Net, GBM (Gradient Boosting machine), RF (Random Forest), IG (Information Gain), and DL-based Multi-Layer Perceptron [MLP]. In addition, more than 100 classification methods were tested on five outcomes: incidence of cartilage volume loss in medial plateau at 48 and 96 months (Prop_CV_48-96), knee replacement (TKR) at 48-96 months, Kellgren-Lawrence (KL) grade ≥2 and medial joint space narrowing (JSN) ≥1 mm at 48 months. Data were retrieved from the Osteoarthritis Initiative and quantitative magnetic resonance imaging (MRI) assessment. This study included 1044-1598 individuals/outcome, as well as 1107 variables and 135 MRI data at baseline. The classification was done using auto-ML tools by calculating the area under curve (AUC) and Matthews correlation coefficient (MCC) for the imbalanced TKR data. To prioritize the selected variables and outcomes, we used the multi-label Sparse Partial Least Square (sPLS) regression method.

Results: Feature selection and sPLS revealed MRI-based variable cartilage thickness and outcome cartilage volume loss at 96 months (Prop_CV_96) as the best predictors of knee OA progression. Moreover, medial joint space width along with pain were among the top predictors. The LASSO method outperformed other feature selection methods (AUC, 0.75-0.91). For the 1-label classification of top common variables between the models, MLP achieved the highest AUC in Prop_CV_96, KL, and JSN (0.80, 0.99, 0.95). For Prop_CV_48 and TKR outcomes, GBM was the best classifier (AUC, 0.70, 0.99).

Conclusion: This is the first time that such a comprehensive study is performed for identifying the best predictors of knee OA rapid progressors. Importantly, data showed that MRI-based variables and outcome have the most significant impact in identifying OA progressors, and could be applied for early prognosis in clinical practice.

« Back to 2019 ACR/ARP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/developing-a-comprehensive-patient-specific-disease-progression-prediction-model-for-knee-osteoarthritis-using-machine-deep-learning-methods/