Prediction of Accelerated Knee Osteoarthritis Using a Convolutional Neural Network

Purpose or Learning Objective: Although knee osteoarthritis (KOA) is typically a slowly progressing disorder, studies have shown that 3% of adults develop radiographic evidence of accelerated KOA (AKOA) over 4 years. Detection of radiographic features and secondary risk factors may provide two main benefits: (1) assist clinicians to identify an at-risk subset of patients with early knee OA, and (2) the ability to identify AKOA, facilitating more precise screening for subjects in clinical trials.

Methods or Background: We formulated the patient selection problem as a multiclass classification task, with classes based on clinically relevant measures of progression (over a time scale typical for clinical trials). Using data from three long-term studies (Osteoarthritis Initiative [OAI], Multi-centre Osteoarthritis Study [MOST], and Cohort Hip and Cohort Knee [CHECK]), we tested multiple algorithms and learning process configurations (including multiclass approaches, cost-sensitive learning, and feature selection) to identify the best performing models. We examined the behavior of the best models, with respect to prediction errors and the impact of used features, to confirm their clinical relevance.

In addition to knee radiographs, body mass index, age, sex, Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) scores, hip symptoms, knee medication injection, and the Kellgren-Lawrence (KL) grading system were used as input for the binary classification models. Selected images included one that had no radiographic knee OA (KL: 0 or 1) at baseline and had high-quality quantitative medial joint space width (JSW) measures on two or more consecutive visits (80,234 unilateral radiographs from all three studies). AKOA was defined as loss of JSW > 10% per 2 years and > 20% per 2 years, to find the best method to predict AKOA. Then OA progression was classified as “slow” or “fast.”

Results or Findings: The XGBoost convolutional neural network (CNN) model achieved the highest performance with an area under the curve (AUC) of 66.16%, trained with the Osteoarthritis Research Society International (OARSI) score of sclerosis and osteophytosis next to the numeric data as input (20% JSW per 2 years). The CNN trained only on the image data resulted in an AUC of 56.26% (10% JSW per 2 years). Combining image and numeric data in a CNN yielded an AUC of 66.3% (20% JSW per 2 years).

Conclusion: The CNN described here shows potential for further development to detect potentially fast-progressing knee OA.

 

MissingLinkText

Die Autoren geben an, dass kein Interessenkonflikt besteht.

Publikationsverlauf

Artikel online veröffentlicht:
02. Juni 2022