Hybrid ST-GCN/HMM Tremor Detector for a Wearable MR-Fluid Exoskeleton

Toufic Jrab

Confirmation: I have read and agree with the IEEE BSN 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Wearable robotics, Tremor suppression, Graph neural networks, Hidden Markov models, Body sensor networks

TL;DR: A 3-block ST-GCN fused with a 2-state HMM detects Parkinsonian tremor in ≤15 ms (AUC 0.70) and is lightweight enough to run on a Jetson Nano for real-time control of an MR-fluid forearm exoskeleton.

Abstract: We present a low-latency tremor-state estimator that couples a three-block spatio-temporal graph convolutional network (ST--GCN) with a two-state hidden Markov model (HMM). Trained on 4\,887 lower-arm IMU windows from 34 Parkinson's disease and control subjects performing activities of daily living (ADLs), the pipeline attains an AUC of~0.70 on held-out subjects and improves negative log-likelihood (NLL) and precision over FFT-threshold, Bayesian, LSTM, and stand-alone ST--GCN baselines. Under an embedded, causal streaming deployment, INT8 inference on a Jetson Nano is \emph{projected} to fit within a sub 80\,ms sensor-to-actuator budget, with ST--GCN compute contributing sub 15\,ms. To our knowledge, this is among the first reports fusing ST--GCN features with probabilistic temporal smoothing for wearable tremor suppression in free-motion ADLs, emphasizing calibrated posteriors for safe actuator triggering.

Track: 3. Signal processing, machine learning, deep learning, and decision-support algorithms for digital and computational health

NominateReviewer: Siamak Ravanbakhsh

Submission Number: 154