Deep-learning model for the prediction of lower-limb joint moments using single inertial measurement unit during different locomotive activities

Abstract

The estimation of lower-limb joint moments during locomotive activities can provide valuable feedback in joint-injury risk evaluation and clinical diagnosis. The use of inertial measurement units (IMUs) in joint moment estimation has drawn considerable attention. Minimizing the number of IMUs deployed on the lower-limb and simplifying the application procedure have long been the pressing problem to be solved during the accurate estimation of multiple joint moments in multiple scenarios. In the present study, the performance of seven different deep-learning models, including the proposed attention-based conventional neural network-bidirectional long short-term memory model and six previous-published typical deep-learning models, were compared in predicting the sagittal plane hip, knee, and ankle joint moments during the most representative locomotive activities. A public dataset was employed to train and validate these models. Seven configurations of IMUs, including different placement and number of IMUs, were evaluated to explore the influence of IMU setup on the performance of the proposed model. Of all seven deep-learning models, best performance was achieved using the proposed model. Pearson correlation coefficient derived from the proposed model using a single IMU attached on the shank, foot, and thigh reached up to 0.85, 0.83 and 0.78, respectively. Shank is the optimal location for attaching a single IMU for the moment prediction of all three joints, while extra IMUs attached elsewhere failed to derive pronounced benefit for improving the estimation accuracy. Thus, even using a single shank-worn IMU, the proposed model is still capable of accurately estimating the joint moments of lower-limbs.