Osteoarthritis (OA) is a common disease which constitutes the fourth leading cause of disability worldwide [1]. According to the US National Health Interview Survey, up to 14 million American people are considered to have a symptomatic knee [2], with additional tens of millions affected as well in Europe, South America, Asia, or Middle East [3]. As a consequence of ensuing healthcare expenditures and losses of activity, the economic burden associated with OA is estimated to represent up to 2.5% of Growth National Product in Western countries [4].
The standard of care for OA based on both non-pharmacological and symptomatic pharmacological treatments has only a limited effect on function and pain. Thus, a very high unmet medical need still persists for a disease-modifying osteoarthritis drug (DMOAD) counteracting disease progression for both function and pain and avoiding the requirement for knee surgical replacement. As of today, the development of such DMOADs has been unsuccessful for two reasons. First of all, significant differences are observed among patients in terms of progression of cartilage degradation. Secondly, in the absence of any established patient stratification in the form of endotypes reflecting well-characterized pathophysiological mechanisms, the slow and heterogeneous evolution of the disease makes it difficult to evaluate the effectiveness of a treatment in a broad patient population, within the 1 or 2 year(s) usual timeframe of a clinical study [5].
In this context, a personalized medicine approach is being considered to treat OA, consisting in identifying the most appropriate target populations predicted to benefit from DMOADs [6]. Primary efficacy endpoints required to document DMOAD efficacy include both clinical variables such as requirement for joint replacement as well as structural changes. The diagnosis of knee OA and the evaluation of its severity are currently based on imaging, with radiography remaining the most commonly used modality in clinical practice [7]. Specifically, knee X-rays are used to determine the JSW (Joint Space Width) as a measurement of the distance between tibia and femur considered as an indicator of cartilage thickness. X-rays of the knee performed for an individual patient at various time points allow to define the JSN (Joint Space Narrowing) as a change in JSW over time [8]. Current regulatory guidelines for clinical trials aiming at evaluating candidate DMOADs recommend that JSN should be used as the primary endpoint in those trials [9].
One limitation, however, is that a reliable evaluation of JSN during patient follow-up remains difficult [10]. A clustering method on OAI data during a 8 year follow-up concluded that only 29% of patients displayed a radiographic progression (as defined by JSN), with no further association between progression and pain worsening [11]. In this context, the use of MRI emerges as a better quantitative endpoint recommended for assessing morphological changes in knee cartilage during OA [12]. MRI allows the assessment of meniscal lesions such as root meniscal tears and extrusions known to be associated with OA progression [13, 14]. It also detects other lesions predictive of pain, such as the presence of synovitis and synovial fluid effusion [15] or bone marrow lesions [16].
We thus undertook the present feasibility study in support of the development of candidate DMOADs with the assumption that the latter should preferably be evaluated in patients likely to progress rapidly. To assess whether knee MR images collected at baseline could predict further cartilage degradation, we implemented a deep learning method using baseline MR images to build up a predictive model for future progression of knee OA, the latter being measured by JSN at 12 months. Additional analyses were conducted in parallel to predict pain grade evaluated by the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC).