From deep learning to transfer learning for the prediction of skeletal muscle forces.

Abstract

Skeletal muscle forces may be estimated using rigid musculoskeletal models and neural networks. Neural network (NN) approach has the advantages of real-time estimation ability and promising prediction accuracy. However, most of the developed NN models are based on conventional feedforward NNs, which do not take dynamic temporal relationships of the muscle force profiles into consideration. The objectives of this present paper are twofold: (1) to develop a recurrent deep neural network (RDNN) incorporating dynamic temporal relationships to estimate skeletal muscle forces from kinematics data during a gait cycle; (2) then to establish a transfer learning strategy to improve the accuracy of muscle force estimation. A long short-term memory (LSTM) model as a RDNN was developed and evaluated. A weight transfer strategy was established. Three databases were established for training and evaluation purposes. The predictions of rectus femoris, soleus, and tibialis anterior forces with developed LSTM network show root mean square error range of 2.4-84.6N. Relative root mean square error (RMSE) deviations for internal and external validations are less than 5% and 10% for all analyzed muscles respectively. Pearson correlation coefficients (R) range of 0.95-0.999 showed perfect waveform similarity between data and predicted muscle forces for all analyzed muscles. The use of weight transfer leads to an improvement of 1.3% for the relative deviation between simulation outcome and LSMT prediction. This present study suggests that the recurrent deep neural network is a powerful and accurate computational tool for the prediction of skeletal muscle forces. Moreover, the coupling between this deep learning approach and a transfer learning strategy leads to improve the prediction accuracy. In future work, this coupling approach will be incorporated into a developed decision support tool for functional rehabilitation with real-time estimation and tracking of skeletal muscle forces. Graphical abstract ᅟ.