| description abstract | Abstract. Essential tremor (ET) is a neurological disorder that causes involuntary rhythmic limb movements, which negatively impact quality of life. ET’s severity is often quantified by The Essential Tremor Rating Assessment Scale (TETRAS). However, this clinical scale suffers from inter- and intrarater error, and still requires expert clinical evaluation. Toward the development of an assessment available outside of the clinic, a long short-term memory (LSTM) network was used to predict ET severity with inertial measurement unit (IMU) data collected from 12 participants with ET rated on TETRAS. LSTM hyperparameters were tuned using a Bayesian optimization algorithm (Optuna library). The LSTM was trained on the normalized data and labels to minimize root mean square error (RMSE) and evaluated using a leave-one-out cross-validation (LOOCV). Across the 12 iterations of the LOOCV, the model achieved an RMSE of 6.72 ± 3.52. The model performs more consistently than naive median guessing but suffers from greater variance; however, the model’s performance improvement over naive median guessing was not statistically significant. Overall, the model’s performance falls close to the 10% interrater error seen in TETRAS. However, additional participants representing a larger range of the TETRAS scale could improve the prediction performance. In summary, this paper presents the first use of TETRAS to train an LSTM to predict ET severity, and motivates the acquisition of large datasets and development of this approach into a flexible app-based severity assessment, which can be administered outside of the clinic. | |
