SR-FSL: Sample reconstruction enhanced few-shot learning for real-time motor unit identification from surface electromyogram.
Yunfei Liu, Zhaoyang Sheng, Dongfang Li and 3 others
PMID 42243970WHAT IT FOUND
Individual motor units can be identified from surface EMG in real time using only 6 seconds of a person's own data per force level, with around 93% agreement with an offline decomposition.
But this was 10 healthy adults, one thumb muscle, no patients.
Key findings
01Across both force levels the proposed method reached an average matching rate of 0.928, which the authors report as over 10% higher than the baseline network, and it significantly outperformed every comparison method.
02The method needed only one 6-second repetition of a subject's own data per force level for fine-tuning, against the 60 seconds of experimental data used by earlier deep learning approaches, but it still required 50 seconds of simulated data for pre-training, so the total data burden of the whole workflow was comparable to previous methods.
03Adding synthetically reconstructed training samples improved identification further, with best performance at a spike shift range of 30 data points and a 1:3 ratio of real to synthetic samples.
STILL TO COME
How it was doneWhat they found
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The method was tested on a single muscle, the abductor pollicis brevis, during thumb abduction. Nothing is known about whether it works on the muscles a therapist would actually want to study in a given patient. All ten participants were healthy adults with no neuromuscular disorder, so the study says nothing about people with stroke, nerve injury, muscle disease or any other condition a therapist treats. The 'ground truth' the method was scored against is itself an offline decomposition algorithm, not a direct measurement of motor unit firing. The 0.928 figure measures agreement with another estimate, and an error in that estimate would not show up. The method still needs 50 seconds of simulated data to pre-train the network, so the total data burden across pre-training and fine-tuning is comparable to earlier approaches. It identifies whether a motor unit fired inside a short window, not the exact time of the spike, which the authors list as future work.
Declared interests
The work was funded by the National Natural Science Foundation of China. The supplied text carries no other funding statement and no competing-interest declaration.
The easy way to misread this
Do not read the roughly 93% agreement figure as evidence this is ready for clinical use. The ten participants were healthy adults with no neuromuscular disorder, recording was from a single thumb muscle in a laboratory, and the comparison standard was an offline decomposition algorithm rather than any patient outcome. No clinical benefit was measured.
Summarised by AI from the full paper, without a clinician reviewing it. Check it against the source before it changes what you do. Read it on PubMed →