A Machine Learning workflow for Diagnosis of Knee Osteoarthritis with a focus on post-hoc explainability

Knee Osteoarthritis (KOA) is a multifactorial disease-causing joint pain, deformity and dysfunction. The aim of this paper is to provide a data mining approach that could identify important risk factors which contribute to the diagnosis of KOA and their impact on model output, with a focus on post-hoc explainability. Data were obtained from the osteoarthritis initiative (OAI) database enrolling people, with nonsymptomatic KOA and symptomatic KOA or being at high risk of developing KOA. The current study considered multidisciplinary data from heterogeneous sources such as questionnaire data, physical activity indexes, self-reported data about joint symptoms, disability and function as well as general health and physical exams' data from individuals with or without KOA from the baseline visit. For the data mining part, a robust feature selection methodology was employed consisting of filter, wrapper and embedded techniques whereas feature ranking was decided on the basis of a majority vote scheme. The validation of the extracted factors was performed in subgroups employing seven well-known classifiers. A 77.88 % classification accuracy was achieved by Logistic Regression on the group of the first forty selected (40) risk factors. We investigated the behavior of the best model, with respect to classification errors and the impact of used features, to confirm their clinical relevance. The interpretation of the model output was performed by SHAP. The results are the basis for the development of easy-to-use diagnostic tools for clinicians for the early detection of KOA.