ºÚÁÏÍø

The title of the thesis: Sensitivity of high-fidelity neural interfaces to perturbations 

Thesis defender: Mansour Taleshi
Opponent: Prof. Utku Åž Yavuz, University of Twente, The Netherlands
Custos: Prof. Ivan Vujaklija, Aalto University School of Electrical Engineering

Robust neural interfaces from muscle signals: what fails, what holds, and how to fix it

This dissertation investigates how reliable non-invasive neural interfaces based on high-density surface EMG and motor-unit decoding are when real-world disturbances occur. The work examines two kinds of perturbations: (1) physiological, where temporary blood-flow restriction (BFR) alters sensory feedback; and (2) technical, where electrical noise, channel loss, and electrode shifts often happen in daily use. Together, these studies reveal where performance breaks, where it is resilient, and which methods can restore robustness.

The main results show that BFR transiently increases discomfort and reshapes the synchronization of motoneurons: alpha-band coherence decreases and low-frequency common drive adapts, while force-tracking accuracy is modestly impaired and then recovers after cuff release. These findings clarify how altered afferent feedback affects the neural drive that EMG-based interfaces rely on. 

On the technical side, widespread (global) electrical noise is the most damaging to motor-unit decomposition and neural-drive estimation, whereas localized noise and moderate random channel loss are far less harmful. For feature-based gesture recognition, amplitude features (RMS/MAV) resist noise and channel loss better than features of signal dynamics; however, electrode displacement remains the toughest challenge and requires explicit mitigation. A synergy-guided channel-clustering approach increases the number of identifiable motor units and improves decoding accuracy, offering a practical path to more stable control.

In sum, the thesis provides new evidence on the sensitivity of HD-EMG neural interfaces to both physiological and environmental stressors and translates those insights into design guidance: preserve signal-to-noise ratio, prefer robust features, account for electrode shift, and use synergy-aware channel selection. These recommendations support more dependable prosthetic, rehabilitation, and human–robot interfaces outside the lab. 

Key words: High-density surface EMG (HD-EMG); motor unit decomposition; neural interfaces; gesture recognition; robustness; blood-flow restriction (BFR); electrical noise; channel loss; electrode shift; muscle synergy

Thesis available for public display 7 days prior to the defence at . 

Contact: mansour.taleshi@gmail.com

Description of the image: Visual summary illustrates how noise degrades the accuracy of motor‑unit‑based motion classification. High‑density surface EMG recordings (Ch001 to Ch192) capture the activity of individual motor units in forearm muscles during an intended wrist flexion task, but the introduction of noise into channels distorts the raw signals. After decomposition, the extracted motor unit spike trains are passed to a machine‑learning classifier that maps firing patterns to predicted movements. The classifier is susceptible to erroneous motion outputs when noise perturbs the input signals. Ref: https://www.biorxiv.org/content/10.1101/2025.09.10.675419v1