Knee osteoarthritis prediction from gait kinematics: Exploring the potential of deep neural networks and transfer learning methods for time series classification.

Recent advances in artificial intelligence methods have allowed improved disease diagnosis using fast and low-cost protocols. The present study explored the potential of different deep neural networks (DNNs) and transfer learning methods to detect knee osteoarthritis patients from gait kinematic time series encoded as image representations. Gait lower limbs kinematic data were collected from 27 patients with knee osteoarthritis and 27 asymptomatic healthy individuals. Joint angles time series were encoded as different image representations (color-based representations, recurrence plots, and Gramian Angular Field). A basic neural network model with three convolutional layers allowed identifying the Gramian Angular Difference Field (GADF) as the best image representation. Then, ten different DNNs with transfer learning and fine-tuning were compared with GADF representations dataset. The best classification performances were achieved with VGG16 and Xception architectures, with an accuracy of 89.7% and 92.3%, respectively. The present study showed the potential of DNNs and transfer learning methods on the development of prediction models related to osteoarthritis, giving insights about time series data classification in health care.