Lower Limb Torque Prediction for Sit-To-Walk Strategies Using Long Short-Term Memory Neural Networks

Joint torque prediction is crucial when investigating biomechanics, evaluating treatments, and designing powered assistive devices. Controllers in assistive technology require reference torque trajectories to set the level of assistance for a patient during rehabilitation or aiding essential daily tasks such as sit-to-walk (STW). STW itself can be generalized into strategies based on individual needs and movement patterns. In this study, three long short-term memory (LSTM) neural networks were empirically trained for hip and knee torque prediction considering these STW strategies and subject anthropometry. The hip and knee are the drivers of STW, while the network architectures were selected for recognizing temporal and spatial relationships. Performance of the LSTMs were compared and evaluated against the STW strategies to accurately generate strategy-specific and user-oriented torque. As such, train and test STW data were obtained from 65 subjects across three age groups: young, middle-aged, and older adults (19-73 years). Model inputs were hip and knee angles with horizontal center of mass velocity, while windowing allowed the LSTMs to dynamically adapt to real-time changes in STW transitions. The encoder-decoder LSTM showcased optimal performance with robust recognition of temporal features. It produced significantly (P<0.05) low hip and knee root mean square error (0.24 ± 0.07 and 0.15 ± 0.02 Nm/kg), strong Spearman's correlation (93.43 ± 2.86 and 84.83 ± 2.96%) and good intraclass correlation coefficients (greater than 0.75), demonstrating model reliability. Hence, this network predicts strategy and user oriented reference torques for more personalized controllers in assistive devices, with natural application of assistance.