Wearable Ultrasound Sensing with Dual Arrays and Machine Learning for Real-Time Tremor Characterization and...

Precise and time-efficient sensing of tremor dynamics is essential for developing personalized, adaptive stimulation strategies for movement disorders such as Parkinson's Disease (PD) and Essential Tremor (ET). Existing modalities-such as inertial measurement units (IMUs) and surface electromyography (EMG)-lack muscle specificity and are prone to stimulation artifacts. To address these limitations, we fabricated and applied a Wearable Ultrasound Sensing (WUS) system, featuring dual flexible 64- element transducer arrays with center frequency of 6 MHz conformed over the flexor carpi radialis (FCR) and extensor carpi radialis (ECR) muscles for simultaneous antagonist muscle monitoring. A machine learning (ML) pipeline was introduced to estimate tissue displacement (TD) from raw ultrasound signals. Across three participants and multiple trials, the WUS system achieved tremor frequency detection within 10 % variation relative to commercial B-mode and IMU sensors, and met statistical equivalence thresholds based on two one-sided test (TOST) analysis. The TD estimation models demonstrated high predictive accuracy, with overall Pearson correlation coefficients between predicted and actual values ranging from 0.88 to 0.97, normalized RMSE values below 10 %, and offline model building times under 800 seconds. Compared to classical B-mode speckle tracking, the proposed WUS-ML pipeline reduced offline model building time by 75 % (from 3095 sec to 792 sec). In real-time performance tests, it reduced frame-level latency from 3700 ms to 17 ms and achieved effective processing rates exceeding 50 Hz, well above the Nyquist threshold for tremor analysis. These results demonstrate the feasibility of dual-array wearable ultrasound for accurate, low-latency tremor tracking and lay the groundwork for future closed-loop afferent stimulation systems in both clinical and at-home environments.