To investigate the longitudinal relationship between physical performance (via real-life accelerometry) and physical capacity (laboratory measurement of gait speed) in patients with knee osteoarthritis (KOA), and to derive accelerometry measured thresholds associated with gait speed decline in KOA that may provide targets for disease-specific physical activity guidelines.
Longitudinal data from the Osteoarthritis Initiative (OAI) accelerometer sub-study was extracted from 1,229 participants assessed 2 years apart. Extracted data include functional capacity, demographic and anthropometric characteristics, patient-reported outcome measures, and accelerometry-based physical activity measures. A “poor capacity” group was defined based on the gait speed quintile decline between baseline and the 2-yr follow-up. A Random Forest classifier was trained to classify individuals’ capacity status, and the impact of each extracted factor on the prediction outcome was analyzed using a novel machine learning interpretation algorithm.
The most impactful predicting feature for gait decline is low minutes in the performance of moderate–vigorous activity (count per min 2,500+). Slower sit-to-stand performance, higher age and self-reported knee pain, and lower minutes in performance light activities (count per min 350-2499) also contributed to the model prediction. The overall classification accuracy is 76.3% (75.4% sensitivity, 76.5% specificity).
We investigated the impact magnitude and direction of each predicting feature on the longitudinal capacity status among KOA patients. Using novel data interpretation method, we established feature thresholds that may increase the probability of gait decline. These identified thresholds may provide meaningful information for establishing specific physical activity guidelines for KOA.