Unsupervised machine-learning algorithms for the identification of clinical phenotypes in the osteoarthritis initiative database

Objectives: Osteoarthritis (OA) is a complex disease comprising diverse underlying patho-mechanisms. To enable the development of effective therapies, segmentation of the heterogenous patient population is critical. This study aimed at identifying such patient clusters using two different machine learning algorithms. Methods: Using the progression and incident cohorts of the Osteoarthritis Initiative (OAI) dataset, deep embedded clustering (DEC) and multiple factor analysis with clustering (MFAC) approaches, including 157 input-variables at baseline, were employed to differentiate specific patient profiles. Results: DEC resulted in 5 and MFAC in 3 distinct patient phenotypes. Both identified a "comorbid" cluster with higher body mass index (BMI), relevant burden of comorbidity and low levels of physical activity. Both methods also identified a younger and physically more active cluster and an elderly cluster with functional limitations, but low disease impact. The additional two clusters identified with DEC were subgroups of the young/physically active and the elderly/physically inactive clusters. Overall pain trajectories over 9 years were stable, only the numeric rating scale (NRS) for pain showed distinct increase, while physical activity decreased in all clusters. Clusters showed different (though non-significant) trajectories of joint space changes over the follow-up period of 8 years. Conclusion: Two different clustering approaches yielded similar patient allocations primarily separating complex "comorbid" patients from healthier subjects, the latter divided in young/physically active vs elderly/physically inactive subjects. The observed association to clinical (pain/physical activity) and structural progression could be helpful for early trial design as strategy to enrich for patients who may specifically benefit from diseasemodifying treatments.