Welcome to MUnitQuest

Aims and scope

MUnitQuest aims to advance methods for reconstructing motor unit spike trains from high-density surface EMG (HDsEMG) through a community-driven competition. The competition will address two distinct algorithmic challenges, yielding two independent leaderboards:

• Isometric challenge. Well-studied, stationary conditions for which multiple neural source separation methods currently exist.
• Dynamic challenge. Less studied, non-stationary conditions in which algorithmic performance remains limited, motivating the development of novel approaches and interactive exchange of ideas.

This is only possible if the competition data includes reliably labeled motor unit spike trains as references. This requires a community-based effort in building a diverse database of experimental data and realistic simulations. We acknowledge this need through an additional challenge:

• Data challenge. Innovative approaches for obtaining high-quality data and labels for the validation and rigorous performance evaluation of neural source separation algorithms (assessed by a panel of experts).

Who is it for

• The Isometric and Dynamic challenges are intended for algorithm developers working on computational methods that can reconstruct the activity of single neurons from complex mixtures.
• The Data challenge is targeted at experimental researchers who use HDsEMG and simulation scientists developing electrophysiological models that want to apply their methods in a highly relevant field of applied neuromuscular research.

How does it work

The competition is organised into three phases (also see the competition timeline):

• Phase 0 (Archive): Community members contribute datasets (i.e., experimental or simulated HDsEMG data together with labeled spike trains), which are automatically standardised to the EMG-BIDS format. Each submitted dataset is assessed via a double-blind review process (also serving as the basis for the awards in the Data challenge). All datasets entering Phase 2 will be released on an open data repository (for datasets not entering the competition’s data collection, this remains optional).
• Phase 1 (Familiarization): In parallel with Phase 0, algorithm developers are provided with training data from the MUniverse benchmark collection (including labels), allowing teams to build, test, and optimize their neural source separation methods (for both algorithm challenges). To qualify for Phase 2, competitors must submit motor unit spike train predictions, which are evaluated through an anonymized temporary leaderboard.
• Phase 2 (Showdown): The main algorithm competition is conducted using the data collection established in Phase 0 (whereby labels are hidden to competitors). Prize-eligible entries are required to share code at the end of the competition; submissions based on proprietary code are not eligible for awards. The final leaderboards for the Isometric challenge and the Dynamic challenge will be based solely on performance during the showdown phase and will be publicly released (non-awarded teams can decide whether their identities remain hidden).

Awards

All teams participating in the algorithmic challenges receive recognition on a permanent leaderboard (per challenge), and all contributions from the Data challenge will be published according to the FAIR (Findable, Accessible, Interoperable, Reusable) principles in an open data repository (mandatory for datasets entering the MUnitQuest data collection, optional otherwise). Furthermore, the top 5 teams per algorithmic challenge (Isometric and Dynamic), as well as the top 5 dataset contributions (Data challenge), will be invited to share their solutions in a special issue of the Journal of Electromyography and Kinesiology.

Motivation and background

Since the development of the concentric needle by Adrian and Bronk nearly 100 years ago, the indirect identification of spinal motor neuron activity from motor unit activity has shaped our understanding of neuromuscular physiology. Over the last 20 years, the development of blind source separation (BSS) algorithms applied to high-density surface electromyography (HDsEMG) recordings has facilitated the study of motor unit activity in living humans, enhanced the population of detectable motor units, and underscored the technique’s potential in applications such as human-machine interfaces. Despite these achievements, several limitations persist:

• Most motor unit identification algorithms are still limited to isometric conditions
• Due to the lack of community-accepted benchmarks, tracking methodological progress is difficult
• The lack of publicly available datasets and standards for sharing EMG data hinders large-scale performance evaluation through automated pipelines

Organizing institutions

• Institute for Modelling and Simulation of Biomechanical Systems, University of Stuttgart
• Department of Bioengineering, Imperial College London

Partners and supporters

• Become a partner! Get in touch (📧) if you want to contribute to our efforts 🙏

Core coordinators

• Thomas Klotz (📧, GitHub)
• Pranav Mamidanna (📧, GitHub)
• Robin Rohlen (📧)
• Oliver Röhrle (📧)
• Dario Farina (📧)

Contributors

• Niklas Enslin
• William Raftery
• Paul Brandenburg