In EMG gesture recognition, the same hand movement can produce different signals when a person changes how forcefully they contract their muscles. For engineers developing myoelectric interfaces, this creates two distinct requirements: a system may need to recognize a gesture regardless of effort, or it may need to distinguish effort levels because they represent different control commands. Hyperdimensional computing offers a way to address both needs within a common classification approach.
This distinction matters in applications such as prosthetic hand control. If a user intends to select a gesture, variations in contraction effort can be an unwanted source of classification error. But if the system uses effort to indicate how firmly to grip, treating every effort level as equivalent would discard useful information.
Why muscle contraction levels affect EMG classification
Surface electromyography (EMG) measures electrical activity associated with muscle contractions using electrodes placed on the skin. Machine-learning classifiers use features derived from these signals to identify gestures. The signals can vary for several reasons, including limb position, electrode movement, and changes in applied force. Some variations are generally undesirable; effort changes, however, may be either a disturbance or an intentional input.
The scale of the challenge can be substantial. The study described in the paper reports prior findings of error rates as high as 50% when a classifier was trained and tested at different force levels, compared with 7% to 19% when training and testing took place at the same level. In its own experiments, classification accuracy dropped by 16.57% to 39.17% across different effort levels, with the largest drop when comparing low and high effort.
These results highlight a practical design issue: a model that performs well under one contraction condition may not transfer reliably to another. A robust system must account for how the training data represents the range of effort expected during use.
How hyperdimensional computing represents gestures
Hyperdimensional (HD) computing represents information using high-dimensional vectors, or hypervectors. The approach combines input information into a representation for each class, then compares new input against stored class representations. In the reported method, the system extracts mean absolute value (MAV) features from EMG signals and encodes information across the electrode channels and over time.
For classification, encoded hypervectors from a gesture class are combined and converted into a bipolar prototype. During inference, the test representation is compared with prototypes in an associative memory using cosine similarity. The system selects the class with the closest prototype.
A key practical feature of the approach is that the study built models using only one of five recorded trials per gesture. This made it possible to test ways of incorporating additional effort contexts without requiring a large set of training examples for every condition.
Two ways to handle effort variation
Recognize the gesture, regardless of effort
When effort should not change the output, the classifier can represent a gesture using data from multiple contraction levels. The paper describes combining representations from different effort contexts to create a gesture prototype that includes more than one level. This can be done by accumulating the encoded data before forming the prototype, or by merging already formed prototypes.
After merging prototypes from two effort levels, the experiments recovered accuracy to above 78.21%. When data from all three effort contexts was accumulated to build a model, accuracy was at least 88.19% across the three contexts. These figures are specific to the study’s dataset and conditions, rather than a guarantee of performance in other systems.
The implication for system design is that training data should reflect the variation the classifier is expected to tolerate. An inclusive gesture model can treat low, medium, and high contractions as contexts of the same intended action instead of separate commands.
Classify both gesture and effort
When effort carries meaning, the model can assign a separate class to each gesture-and-effort combination. For example, a gesture performed at different contraction levels becomes multiple output classes rather than one. This lets the system distinguish an action from its effort level, a useful distinction when a control scheme needs to represent different levels of force.
The trade-off is that the number of classes increases. The paper also discusses using multiple prototypes for one gesture—one for each effort level—while retaining the gesture label. This can improve gesture classification compared with merging prototypes, but requires storing and comparing more prototypes.
What the experiment tested
The study collected EMG data from five able-bodied adult male subjects performing nine gestures at low, medium, and high effort. The gesture set included individual finger flexions and extensions, along with multi-finger hand postures. A 64-channel electrode array recorded forearm activity at 1 kS/s. Participants received visual feedback based on mean signal energy across channels and were asked to target 25%, 50%, and 75% of their maximum voluntary contraction.
Each gesture trial included transition periods and a steady hold period. The classifier used MAV features calculated over non-overlapping 50-sample windows, with spatiotemporal encoding across channels and time. These details define the experimental context and are important when interpreting the results: performance may depend on the sensing setup, gesture set, and data collection conditions.
Practical implications for automation and assistive control
For developers of EMG-based interfaces, the first decision is functional: should effort variation be ignored, or should it be exposed as part of the command? That choice shapes the class labels, training data, and memory requirements. A gesture-only model should include representative effort contexts so that changing force does not unnecessarily change the recognized action. A gesture-plus-effort model should define effort levels as distinct outputs and account for the larger class set.
Hyperdimensional computing provides a flexible framework for both choices. In the reported experiments, adding effort contexts or defining additional classes did not require a fundamentally different classification pipeline. However, the results should be evaluated against the intended application and its own users and operating conditions before being treated as evidence of deployment performance.
Conclusion
Muscle contraction level is not simply noise in EMG gesture recognition: depending on the application, it can be a variation to overcome or a control signal to preserve. The study shows how hyperdimensional computing can support both effort-robust gesture recognition and classification of gesture-and-effort combinations. For technical teams, the central design task is to decide what effort means in the interface, then build the model and training data to match that requirement.