Where are my electrodes? Quantifying high-density surface electromyographic array shift through pairwise correlation mapping

Introduction:: Traditionally, post-stroke patient rehabilitation progress is evaluated during in-person functional assessment sessions, but recently remote assessments, held using video-conferencing platforms, have become commonplace due to their accessibility [1]. This reliance on video-based observational data has resulted in greater subjectivity when drawing conclusions on patient progress, and so there is a demand for more quantitative evidence to support patient recommendations.

Despite surface electromyography (sEMG) demonstrating potential in quantifying stroke recovery at a neuromuscular level, the technology’s repeatability deficits due to placement variations between recording sessions restrict the conclusions that might be drawn from the data over time. As such, organisations like SENIAM have gone a long way towards standardising the anatomical landmarks used to inform sEMG sensor placement, but even slight variations in this can dramatically alter the features extracted from the data, making the technology ill-suited for at-home use in non-expert populations.

In response, this study hypothesises that 64-channel high-density sEMG (HDsEMG) arrays can be used to generate neurophysiological maps of lower-limb muscle activity, and that by aligning these across recording sessions we can quantify how the array has moved relative to its original position. By confirming this hypothesis, we hope to provide researchers and clinicians a means of extracting sEMG data from a consistent location across sessions, something that we hope will improve the confidence with which variations in signal properties can be attributed to neurophysiological changes in patient.

Materials and Methods:: To explore our hypothesis, we recruited 3 healthy individuals under ethics approval H22-02423 to perform a selection of isometric and common stroke assessment exercises while instrumented with HDsEMG arrays on the tibialis anterior and gastrocnemius medialis. We used an 8x8 1cm-spaced electrode grid (OT Bioelettronica HD10MM0808) connected to a Sessantaquattro+ (OT Bioelettronica) wireless amplifier, and each exercise set consisted of 3x10 second isometric contractions, 30 seconds of sit-to-stand repetitions, and a 10-pace timed-up-and-go. To record array location after each exercise set, a 3D scan (Structure Sensor Pro and Apple iPad Pro 5th-Generation) was taken of the leg with fiducials located on the array itself as well as anatomical landmarks at the knee and ankle. From here, the array was shifted randomly within set bounds (|translation| ≤4.25cm, |rotation| ≤30o) and the aforementioned exercise was repeated; this continued until there were four sets of exercise and scan data for each muscle with different HDsEMG placements.

To extract the ground truth shift values, post-shift scans were registered to those of the original location using knee and ankle anatomical fiducials. Then, the array fiducial and corner locations were projected onto a 2D coordinate axis with the transform mapping the original array to that post-shift computed. For the HDsEMG data, we computed the pairwise correlation between electrodes following standard pre-processing, and spline interpolated to generate a high-resolution map of the underlying musculature. We then performed a global search to identify the transform that minimized the error between the high-resolution array correlation maps pre and post-shift.

Results, Conclusions, and Discussions:: Preliminary analysis on a single participant’s data has shown the potential of the HDsEMG algorithm to identify array shift between exercise bouts. When the array was shifted by -0.55cm laterally, 1.04cm superiorly, and rotated 10.07o counter-clockwise about its center, the HDsEMG algorithm predicted a -0.49cm, 1.12cm, and 10.82o shift when using gastrocnemius medialis isometric contraction data. This result meant that the maximum Euclidean distance between any electrodes in the ground truth and those extracted from the HDsEMG data was 0.16cm, a value less than a fifth of the array’s 1cm inter-electrode distance.

While data has, thus far, only been analyzed for one participant and collected from a further two, we intend to collect data from at least 15 healthy participants to ensure the algorithm’s generalizability. Furthermore, our initial analysis was performed solely on isometric contraction data given its high signal-to-noise ratio and lack of motion artefacts, but with time we hope to demonstrate the algorithm’s practical utility by extending it to the more dynamic stroke assessment exercises.

With regards to the 3D scans, the ground truth shift values rely on several key assumptions affecting accuracy. First, the 1in-diameter fiducial locations are manually selected, resulting in an error of approximately 3mm when identified in the same scan on multiple occasions. Additionally, after identifying the array’s corners and fiducial points, these were projected onto a 2D plane to compute the shift, not accounting for the contours and curvatures of the body. Because the shift magnitudes are < 4.25cm, we expect this projection to produce an error of around 0.15cm when approximating the leg as a 5cm-radius cylinder. When compounded, these errors are of the same order of magnitude as the shift difference seen between the HDsEMG method and 3D scans, so moving forward we will use edge detection methods to determine the fiducial centre locations and trace the array shift over the skin’s surface to reduce inaccuracies.

Regardless, we have demonstrated that our method can identify the relative displacement of an HDsEMG array across recording sessions which can provide more consistent and robust muscle activity measurements in clinical and longitudinal care settings.

Acknowledgements (Optional): :

References (Optional): : [1] Eng JJ, Pastva AM. Advances in remote monitoring for stroke recovery. Stroke. 2022 Jul 1; 53(8):2658–61.