Abstract P402: Lifestyle Management Empowered by Deep Transfer Learning of Biomechanical Dynamics

Lifestyle management is essential for many promising applications, like cardiac health monitoring and post-stroke rehabilitation. Biomechanical dynamics during human physical movements are therefore attracting more attentions towards real-time and long-term lifestyle management. However, one challenge in biomechanical dynamics analysis is how to detect diverse physical activity types. Even more challenging is, how to quickly tune a detection algorithm that can adapt easily to a new user, who has user-specific biomechanical dynamics. We proposed and developed a deep transfer learning algorithm that, not only detects the physical activity types with a deep neural network, but also transfers the similar knowledge from non-target users to the target user. The former effort enables the algorithm to analyze the complex biomechanical patterns effectively, and the latter effort minimizes the tuning time of the algorithm when it is used on a new user. Evaluated on a fifteen-subject database, with six physical activity types, the proposed deep transfer learning algorithm achieves a detection accuracy of 90%, 92%, and 94%, with 0%, 10%, and 50% of training data from the target user, respectively. The results demonstrate that, with the transferred knowledge only, the accuracy can already reach 90%, which can be further increased with additional turning effort using the target data. This study will greatly advance deep learning of biomechanical dynamics, and further the transfer learning approach for algorithm tuning effort minimization on the new user. This study will greatly benefit lifestyle management by leveraging deep intelligence.