Deep-learning models classified short sensor windows with accuracy close to 99% on a new public dataset combining muscle and movement signals. The strongest representative model, a CNN-LSTM with fine-tuned attention, reached 99.2% accuracy, alongside 99.1% precision, 99.0% recall and a 99.0% F1-score on the study’s test subset.
The dataset, called SDALLE, pairs surface electromyography, or EMG, with inertial measurement unit, or IMU, signals. In practical terms, it brings together muscle-activity and motion readings from nine healthy male participants aged 12 to 39. The recordings cover walking, jogging, stair ascent and stair descent, and the researchers prepared a harmonized comparison with the public ENABL3S dataset.
The exported dataset has 56 input channels from eight sensors. Three magnetometer channels were excluded, leaving one EMG channel and six inertial channels for each sensor.
A strong score with a narrow test
The headline result is a window-level score. The continuous recordings were divided into short segments, each containing 16 time steps. The stride was four time steps, so adjacent windows overlapped by 75% and shared 12 of their 16 samples.
Before the split, feature-wise z-score standardization, a way of putting measurements on a common scale, was fitted using the complete merged dataset. The windows were then shuffled into 70% training and 30% testing, with seed 42, and no participant or trial grouping.
That setup makes the benchmark subject-dependent: windows from the same participants, and potentially the same trials, could appear in both subsets. The article says the resulting accuracy values may be optimistic and do not estimate performance on people the models have not seen.
Under the implemented window-level protocol, attention-enhanced CNN-LSTM models recorded about 98% to 99% accuracy on ENABL3S and about 99% on SDALLE. All architectures were trained with Adam for 30 epochs, using batches of 64 and a learning rate of 0.001.
More complex models took longer to train
To align the two sensor streams for joint analysis, EMG was recorded at 1,259 Hz, while the IMU signals were sampled at 148 Hz. The EMG was then downsampled to 148 Hz by cubic interpolation.
The comparison also exposed a practical tradeoff. Hybrid and attention-enhanced models took longer to train, while inference time remained relatively low on the study workstation. The measurements were specific to the hardware and processing pipeline and did not validate deployment on wearable devices.
Synthetic data did not raise the average score
The study included an exploratory test of WGAN-GP, a generative model used to create additional time-series rows for the less common activities. For each of the nine participants, the researchers added 2,000 synthetic rows for jogging, stair ascent and stair descent. The dataset grew from 446,861 to 500,861 rows, an increase of about 12%, while walking’s share fell from 36.3% to 32.4%.
Across 16 model variants, average testing accuracy fell descriptively from 96.9% before augmentation to 96.5% afterward, a difference of 0.4 percentage points. Average precision moved from 96.7% to 96.3%, recall from 96.6% to 96.4%, and F1-score from 96.7% to 96.3%.
The researchers treated this as a descriptive comparison. It did not include class-wise bias analysis, repeated-run variation, confidence intervals, effect sizes or inferential statistical tests, so it does not establish that augmentation reduced majority-class bias or improved prediction stability, robustness or generalization. The GAN result was presented as exploratory class-balancing behavior.
A dataset for further testing
SDALLE’s raw recordings, processed datasets and preprocessing scripts are publicly available on Zenodo. The release gives other researchers access to the materials needed to examine the benchmark under different validation designs.
The evidence is confined to nine healthy male participants and the controlled, four-activity recordings used here. It does not establish performance on unseen participants or real-time operation on wearable hardware. Participant-level validation and device testing are needed to answer those questions.
The article reports that it was received on 28 February 2026 and accepted on 28 July 2026. Open-access funding came from the Science, Technology & Innovation Funding Authority in cooperation with the Egyptian Knowledge Bank.
Paper data and sources
Original title: A multimodal EMG-IMU dataset and multi-dataset benchmark for deep learning-based human activity recognition.
Authors: Mohamed Khaled Farouk, Mohamed Fawzy El-Khatib, Mohammed I Awad et al.
Journal/Repository: Scientific reports
Status: Peer-reviewed
First online: 2026-08-20
DOI: 10.1038/s41598-026-64847-4
Original paper · Full text