Authors (including presenting author) :
Li HHT (1)(2), Chan LC (2), Chan PK (3), Wen C (2)
Affiliation :
(1) Department of Prosthetics and Orthotics, Tuen Mun Hospital,
(2) Department of Biomedical Engineering, The Hong Kong Polytechnic University,
(3) Department of Orthopaedics and Traumatology, Queen Mary Hospital, The University of Hong Kong
Introduction :
Knee Osteoarthritis (KOA) is amenable to prevention and early treatment, but the waiting time for new case booking at O&T SOPCs is 18 weeks on average and 108 weeks longest. Accurate patient stratification based on progression risk can provide optimal disease management and treatment outcomes. However, there lacks a reliable and efficient prognostic tool for mass screening, leading to delayed diagnosis.
Objectives :
(1) To develop an interpretable machine learning (ML) model for KOA prognosis that benefits low-cost screening and triage.
(2) To identify important risk factors for predicting KOA progression.
Methodology :
(1) The Osteoarthritis Initiative (OAI) dataset was used to extract the 15 types of easily accessible clinical data, e.g. patient demographics, modifiable lifestyle-related outcomes, co-morbidities, lower limb conditions, and self-reported outcomes, for analysis.
(2) The optimal KOA progression was defined as the simultaneous progressions of Kellgren-Lawrence (KL) Grade (i.e. any increase in KL Grade, except from Grade 0 to 1), and WOMAC Pain score by at least 1.8 points within 4 years after the first clinical visit, from our previous study.
(3) A total of 2,179 knee samples were used to develop the ML model. The Decision Tree was employed with the Self-paced Ensemble to account for imbalanced classification and a train-test split of 8:2 for independent testing and cross-validation.
(4) Feature importance (i.e. risk factor contribution) was visualised by Shapley Additive Explanations (SHAP).
Result & Outcome :
Results:
(1) The ML-driven KOA prognostic model performed with the AUC score of 0.784 ± 0.018 from the OAI dataset, similar to other models that used MRI as predictors, which is a gold standard for early identification of KOA.
(2) The SHAP analysis revealed WOMAC Disability Score, BMI, the use of pain medication, history of knee surgery, and age are the most influential prognostic predictors.
Conclusions:
(1) Our big data analytic approach with real-world data for patient-specific KOA prognosis was the first study in Asia.
(2) The use of simple and easily accessible predictors could expedite low-cost yet reliable community screenings.
(3) The interpretable feature importance analysis by SHAP can promote patient self-management.
