Auto-FES: Clinical Evaluation of Automatic FES
The AUTO-FES project evaluates a data-driven, unsupervised calibration framework designed to automate the setup of multi-array functional electrical stimulation (FES) systems. Built upon a robust Bayesian optimisation algorithm, the technology models user-specific neuromuscular responses to rapidly identify optimal stimulation parameters from sparse data, significantly reducing setup times. To transition this framework from controlled lab validation in healthy cohorts to real-world clinical applications, the project collects and analyses kinematic and physiological data from post-stroke patients. By incorporating patient-specific characteristics, such as spasticity, abnormal muscle synergy, and altered muscle tone, into a hybrid probabilistic model, AUTO-FES adapts the system for clinical environments. The initiative aims to advance the maturity of the technology, establishing a robust, easy-to-operate assistive solution suitable for unsupervised use in clinical and home settings.
Motivation
Persistent upper-limb paretic symptoms affect the majority of stroke survivors, severely limiting their independence in rehabilitation and activities of daily living. While multi-array FES is a highly promising technology for assisting and restoring complex hand movements, its clinical and commercial adoption is bottlenecked by the requirement for manual configuration at each use. Identifying effective stimulation sites via traditional trial-and-error methods is time-consuming, causes early muscle fatigue, and relies heavily on clinical expertise. This structural dependency restricts high-tier therapy to supervised clinical environments, preventing patients from achieving the high dose and frequency of training required for functional neural recovery at home.
Research at ITR
Our research method centres on a data-driven, active learning framework designed to automate the calibration of multi-array functional electrical stimulation (FES) systems.
The core methodological components include:
- Probabilistic Neuromuscular Modelling: We model the highly variable, non-linear stimulation-response space using a hybrid probabilistic framework to capture patient-specific muscle characteristics.
- Sample-Efficient Bayesian Optimisation: Instead of using exhaustive, time-consuming sequential searches across electrode pairs, the system employs an active search algorithm to efficiently explore the stimulation space. This identifies optimal stimulation configurations within a strict trial budget to avoid early muscle fatigue.
- Cross-Session Refinement: Rather than starting from a neutral state at every application, the algorithm inherits the posterior data from previous sessions as a baseline prior. This progressive refinement should significantly accelerates the setup process across multiple clinical visits.
The clinical utility of the calibrated hand configurations is subsequently validated using standardised object manipulation assessments.
Related research topics:
Team Member
- Sandra Hirche (principal investigator)
- Satoshi Endo
- Hossein Kaviani rad
- Nicolas Kantor Nagel