Muscle activation prediction in essential tremor through neuromusculoskeletal digital twinning and deep neural networks

Essential tremor (ET) is the most prevalent movement disorder among adults, affecting up to 5% of the population over 65 years of age. Accurate, individual-leven prediction of ET dynamics is crucial for optimizing therapies, such as sub-motor threshold stimulation (delivering electrical currents below the threshold for motoneuron activation), where the timing of stimulation is key for effective tremor reduction. Although there have been some efforts to implement machine learning predictive models, real-time prediction and estimation of muscle activation are still challenging due to the closed-loop nature of neuromuscular control, sensor noise, signal transmission delays, and scarcity of data. In this study, we develop and evaluate a digital twin model to train neural networks for real-time muscle activation prediction in ET. We describe how a digital twin of ET, a computational neuromusculoskeletal model of ET deployed in the SCONE simulator, allows for properly training deep recurrent neural networks (RNN) to predict muscle activation. Moreover, it enables parameterized synthetic simulation of tremor. Results on predicting muscle activation from wrist flexo-extension movement show that the RNN has an average prediction accuracy of 81% and 83% with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) gated neurons, respectively. While this work still uses only synthetic data, it shows the potential for treatment optimization and personalized therapeutic strategies, such as peripheral electrical stimulation.