⚡ electrical engineering

Improving the clinical utility of lower-limb surface electromyography (sEMG) by quantifying and correcting for location changes in inter-session recordings

This study presents a novel high-density sEMG algorithm that quantifies and corrects for inter-session electrode location shifts in lower-limb muscles, thereby significantly reducing feature variability and enhancing the clinical utility of sEMG for tracking recovery in neurological and musculoskeletal disorders.

Original authors: Fraser Douglas, Mona Pei, Quoc Sy Vu, Linh Le, Calvin Kuo

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fraser Douglas, Mona Pei, Quoc Sy Vu, Linh Le, Calvin Kuo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Moving Target"

Imagine you are trying to take a perfect photo of a specific flower in a garden to track how it grows over the next few weeks. To get a fair comparison, you need to take the photo from the exact same spot every time.

However, if you accidentally step two feet to the left on day two, the flower looks different in the photo—not because the flower changed, but because your viewpoint changed.

This is exactly the problem with sEMG (surface electromyography), a technology used to listen to muscle activity. Doctors and researchers place sticky sensors on a patient's skin to "listen" to muscles. If they place the sensors in a slightly different spot during a second visit (even by just a centimeter), the signal changes. It's like the flower photo again: the data looks different, but it's just because the "camera" moved, not because the patient's muscle got better or worse. This makes it very hard to track recovery over time, especially for patients who can't help place the sensors perfectly (like stroke survivors) or when patients do the exercises at home.

The Solution: A "Smart Map" Algorithm

The researchers developed a new computer program (an algorithm) that acts like a GPS for muscles.

Instead of just assuming the sensors stayed in the same spot, this program looks at the "sound" of the muscle activity from a high-density grid of 64 tiny sensors (like a high-resolution camera sensor). It compares the "sound" from the first session to the "sound" from the second session.

By analyzing how the patterns of muscle activity shifted across the grid, the program can mathematically calculate: "Ah, the sensors moved 2 centimeters to the left and rotated slightly."

How They Tested It

To prove this worked, they didn't just guess; they set up a controlled experiment:

The Setup: They put the sensor grid on the legs of 11 healthy volunteers.
The "Ground Truth": They took a precise 3D scan of the leg with the sensors on it. This is like taking a perfect, high-definition map of where the sensors actually were.
The Move: They took the sensors off, moved them intentionally to a new spot (shifting them up to 4 cm and rotating them), and put them back on.
The Test: They asked the computer program to guess where the sensors moved, without looking at the 3D scan. Then, they compared the computer's guess to the actual 3D map.

The Results: Getting Closer to the Truth

The results showed that the computer program was much better at finding the new location than just assuming the sensors hadn't moved at all.

The "No-Shift" Guess: If you assume the sensors didn't move, you are wrong about 82% of the time.
The "Smart Map" Guess: The algorithm correctly identified the new location in 82% of cases.
Precision: In about 38% of cases, the algorithm pinpointed the new location within 1 centimeter (about the width of a fingernail) of the actual spot. This is impressive because even trained medical professionals often struggle to place sensors that accurately on patients with limited muscle control.

Does It Fix the Data?

The ultimate goal was to see if fixing the location guess actually made the muscle data more consistent.

Without the fix: If you compare the same sensor number from Session 1 to Session 2 (even though it's now over a different part of the muscle), the data looks very different (a 21% difference in signal strength).
With the fix: The algorithm finds the "closest match." It says, "Sensor #10 in Session 1 is actually over the same muscle spot as Sensor #12 in Session 2." When they compared these matched spots, the data looked almost identical to the "perfect" 3D scan comparison.

The Bottom Line

This paper shows that we can use a smart computer algorithm to figure out exactly where muscle sensors moved between visits. By doing this, we can "correct" the data so that we are comparing apples to apples, rather than apples to oranges.

This means that in the future, we might be able to track muscle recovery more reliably, even if the sensors aren't placed in the exact same spot every time, making it easier to monitor patients' progress over weeks or months.

Technical Summary: Improving the Clinical Utility of Lower-Limb sEMG via Location Quantification

Problem Statement
Surface electromyography (sEMG) offers a non-invasive window into muscle activity, holding significant promise for assessing neurological and musculoskeletal recovery, particularly in stroke rehabilitation. However, its clinical adoption is hindered by high sensitivity to electrode placement. Consistent sensor placement is critical to ensure that signal changes reflect physiological recovery rather than artifacts of sensor movement. While guidelines like SENIAM exist, they rely on palpation and anatomical landmarks, a process that is difficult for impaired populations (who may lack muscle tone) and non-expert users (e.g., in remote tele-rehabilitation settings). Previous attempts to mitigate this using High-Density sEMG (HDsEMG) and Independent Component Analysis (ICA) were limited to small, translational shifts in specific directions and did not account for rotational variations or larger placement errors common in clinical practice.

Methodology
The study developed a novel algorithm to quantify and correct for inter-session electrode location changes using HDsEMG data from 11 healthy participants.

Data Collection: Participants performed isometric (maximal contractions) and dynamic exercises (Sit-to-Stand, Timed-Up-and-Go) with 64-channel HDsEMG arrays (8x8, 1cm spacing) on four lower-limb muscles: gastrocnemius medialis (GM), tibialis anterior (TA), semitendinosus (ST), and tensor fascia latae (TFL).
Ground Truth Generation: To simulate clinical imprecision, arrays were reapplied at shifted locations (up to 4.25cm translation and ±30° rotation) across four segments. A 3D scanner captured the array's position relative to skin markers to establish "ground truth" shifts ( $x, y, \theta$ ).
Algorithm Development: A gradient-descent-based global optimizer was designed to estimate the shift between recording sessions. The algorithm minimized a weighted objective function based on the Root Mean Squared Error (RMSE) of three spatially mapped metrics extracted from the HDsEMG signals:
1. Mean absolute pairwise correlation (spatial pattern of muscle activity).
2. Mean frequency (spectral content).
3. Maximum envelope amplitude (time-domain magnitude).
  The metrics were normalized, and weights were determined (2.72 for correlation, 1.24 for mean frequency, 1.00 for amplitude) to balance their contributions.
Validation: The algorithm's estimated shift was compared against the 3D scan ground truth using Mean Euclidean Distance (μED). Furthermore, the study analyzed whether using the algorithm to identify "closest" electrodes across sessions reduced feature variability compared to using the "same" electrodes (the current standard practice).

Key Contributions

Robust Shift Estimation: The paper presents an algorithm capable of estimating both translational and rotational electrode shifts using only sEMG signal features, without requiring external imaging during the actual recording session.
Feature Mitigation Strategy: It demonstrates a method to mitigate inter-session variability by mapping features to the closest electrodes identified by the algorithm, rather than relying on fixed electrode indices.
Comprehensive Validation: The approach was validated against 3D scan ground truth across multiple muscles and both isometric and dynamic exercise conditions, providing a realistic assessment of performance under clinically relevant placement errors.

Results

Location Estimation: The algorithm improved the location estimate over the assumption of no shift (origin assumption) in 81.7% of isometric cases. The median μED was reduced from 3.00 cm (origin assumption) to 1.22 cm (algorithm). Specifically, 37.6% of all shifts were identified within 1 cm (the inter-electrode distance), compared to less than 5% for the origin assumption. Performance varied by muscle, with the Tibialis Anterior (TA) showing the highest accuracy (50.0% within 1 cm).
Feature Variability: For isometric exercises, feature differences (relative percent difference and mean-normalized standard deviation) computed between "closest" electrodes (via algorithm or ground truth) were statistically similar to each other but significantly lower than differences computed using the "same" electrodes across shifts. For example, the mean difference for isometric max envelope amplitude increased from 15.9% (with algorithm correction) to 21.1% (without correction).
Dynamic Exercises: While the algorithm showed slightly lower success rates for dynamic exercises (improving estimates in ~56–63% of cases and identifying shifts within 1 cm in ~15–17% of cases), the trend of reduced feature variability for the "closest" electrode approach remained consistent for most features, though statistical significance was less pronounced than in isometric data.

Significance and Claims
The authors claim this work represents a "promising step toward robust sEMG measurement" for tracking musculoskeletal and neurological recovery. By effectively reducing inter-session feature differences arising from electrode location changes, the proposed approach facilitates more direct cross-session feature comparisons. The paper suggests that this algorithm could enable longitudinal sEMG assessments where signal changes are attributed to physiological recovery rather than sensor placement errors.

The authors remain modest regarding limitations, noting that:

The study was conducted within a single 3-hour session, and multi-week generalizability (critical for stroke recovery tracking) requires further validation.
The population was healthy, not stroke patients, though the TA results during dynamic exercise are highlighted as particularly relevant for drop-foot assessment.
The algorithm relies on global optimization, which may encounter local minima, and performance is bounded by the requirement for some physical overlap between array positions.

Ultimately, the paper positions this technology as a solution to the "expert knowledge" barrier in sEMG setup, potentially enabling objective recovery quantification in remote or non-expert settings.