Document Type : SYSTEMATIC REVIEW
Department of Orthopedics, Mayo Clinic, 200 First St SW Rochester, MN 55905
1Department of Health Sciences Research, Mayo Clinic, 200 First St SW Rochester, MN 55905
Department of Health Sciences Research, Mayo Clinic, 200 First St SW Rochester, MN 55905
Introduction Knee osteoarthritis (OA) is a prevalent joint disease. Clinical prediction models consider a wide range of risk factors for knee OA. The aim of this review was to evaluate published prediction models for knee OA and identify opportunities for future model development.
Methods We searched Scopus, PubMed and Google Scholar using the terms knee osteoarthritis, prediction model, deep learning and machine learning. All the identified articles were reviewed by one of the authors and we recorded information on methodological characteristics and findings. We only included articles that were published after 2000 and reported a knee OA incidence or progression prediction model.
Results We identified 26 models of which 16 employed traditional regression-based models and 10 machine learning models. Four traditional and 5 machine learning models relied on data from the Osteoarthritis Initiative. There was significant variation in the number and type of the risk factors. The median sample size for traditional and machine learning models was 780 and 295, respectively. The reported Area Under the Curve (AUC) ranged between 0.6 and 1.0. Regarding external validation, 6 of the 16 traditional models and only one of the 10 machine learning models validated their results in an external data set.
Conclusion Diverse use of knee OA risk factors, small, non-representative cohorts and use of magnetic resonance imaging which is not a routine evaluation tool of knee OA in daily clinical practice are some of the main limitations of current knee OA prediction models.