Classification of Upper Limb Movements based on a LSTM Model in Aquatic Rehabilitation

The objectives of this study were to improve aquatic rehabilitation through the utilization of LSTM model based activity classification by (1) identifying distinct combination of time and frequency domain features of sEMG using correlation analysis, (2) determining the optimal learning rate (0.001) for all parameters (sensitivity, specificity, precision, and accuracy) of the LSTM network model in both on-land and aquatic environments and (3) showcasing the subject-specific accuracy of the LSTM network model with a learning rate of 0.001. To obtain the datasets for the feature extraction, ten healthy male subjects performed bicep curls (BC) and tricep kickbacks (TK) in both on-land and aquatic environments. The research findings indicated that optimizing the sensitivity (87.4% and 86.8%), specificity (90.1% and 89.8%), precision (89.1% and 88.8%), and accuracy (91.4% and 90.8%) for on-land and aquatic environments respectively of the LSTM network required a learning rate of 0.001. The LSTM model eventually exhibited a commendable level of accuracy in classifying BC across conditions of land (90.21% to 93.65%) and aquatic (90.1%), albeit with a marginal decline. Additionally, the wide range of accuracies in TK (88.56% to 92.69%) demonstrated the capability of LSTM networks to generate more precise and comprehensive evaluations of muscular movements, thereby emphasizing their potential in the development of aquatic rehabilitation programs that are both individualized and efficient.Clinical Relevance—This study and relevant experiment demonstrate improved accuracy of the LSTM model discerning muscle movements of the upper limbs, underscoring the potential of the model to assist clinicians and therapists in developing individualized and efficacious aquatic rehabilitation programs.