Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap

Synchronized AMG and EMG Dataset of Lower-limb Muscle Activities in Everyday Training

This publicly released dataset contains time-aligned acceleromyography (AMG), surface electromyography (EMG), optical motion capture (MoCap), and four knee/ankle joint-angle signals from 30 subjects. The repository also provides the frozen 5–100-Hz benchmark code, 64 fitted model artifacts, fixed reference predictions, source tables, and integrity manifests used for the associated dataset paper.

Repository: https://huggingface.co/datasets/Tdongxu/A_Synchronized_Lower_Limb_AMG_sEMG_and_Mocap

Persistent identifier: https://doi.org/10.57967/hf/9950

Repository version: 2.0.2

Version 2.0.2 is a Dataset Viewer configuration patch over version 2.0.1. It defines nine independent tabular configurations so heterogeneous CSV files are not inferred as one split. The repository slug, trial NPZ keys, window-index filenames, window IDs, model weights, predictions, metrics, and figure source values remain unchanged. See RELEASE_NOTES_v2.0.2.md for the patch scope and metadata/naming_compatibility_v2.csv for the display-name mapping introduced in version 2.0.1.

Dataset Viewer configurations

The Dataset Viewer exposes nine intentionally separate tables: trials (default), subjects, conditions, channels, auditory_cues, excluded_trials, primary_benchmark, benchmark_window_counts, and model_inventory. Each configuration points to exactly one CSV file with its own schema. Other CSV, CSV.GZ, NPZ, model, prediction, metric, and source-data files remain downloadable repository assets but are not automatically merged into the default Viewer split.

Dataset scope

• 30 de-identified subjects (user01–user30);
• 11 functional lower-limb actions and treadmill walking at 1–5 km/h;
• four planned repetitions for each of 16 conditions;
• 1,918 released trials from 1,920 expected recordings;
• 16 triaxial accelerometers, with four sensors at each of four muscle sites (48 AMG axes);
• four matched EMG channels;
• 15 named MoCap markers;
• four joint-angle outputs in the canonical order right knee, left knee, right ankle, and left ankle;
• cross-subject train/validation/test splits of 20/3/7 subjects;
• within-subject repetition splits using A1+A2/A3/A4 for train/validation/test.

Two user23 trials—2-km/h treadmill walking repetitions A3 and A4—were excluded because a required AMG sensor cluster was missing. The exclusions are listed in metadata/excluded_trials.csv.

Acquisition clocks and time alignment

Acquisition used a Windows host and four Raspberry Pi units whose system clocks were coordinated by NTP. A central program initiated acquisition, but the four accelerometer pods wrote independent files with recorded file-start timestamps. There was no shared hardware trigger.

For processing, each accelerometer time axis was constructed from its recorded file-start timestamp and the nominal 1600-Hz interval. The 48 axes were linearly interpolated without extrapolation onto their common nominal 1600-Hz grid. EMG remained at its native 1000 Hz, and MoCap retained native frame timestamps. The modalities were cropped to common support. No empirical hardware-offset correction or clock-drift correction was applied. Auditory cue records were not used for alignment.

The nominal 1600-Hz processing grid has an 800-Hz Nyquist frequency. This is not a claim of a confirmed 800-Hz effective sensor bandwidth. Accelerometer values are published in recorded device units because register-dependent physical sensitivity was not established by the available acquisition record.

Trial files

The 1,918 NPZ files under data/trials/ are unchanged from repository version 1.0.0. Each file contains:

• amg_raw_aligned
• amg_dc_corrected_aligned
• amg_highpass_5hz_aligned
• semg_raw
• semg_preprocessed_basic
• motion_markers_named
• motion_skeleton_selected
• joint_angles_deg
• joint_angle_valid_mask
• time_amg_s, time_semg_s, and time_motion_s
• metadata_json

The three time arrays are float64 seconds relative to the released shared trial origin. Absolute acquisition timestamps and dates are not published. See metadata/npz_data_dictionary.csv for shapes, dtypes, and units.

The semg_* and time_motion_s keys are stable machine identifiers from the published schema. They are retained to keep existing loaders and citations working; the reader-facing names introduced in version 2.0.1 are EMG and MoCap.

Primary 5–100-Hz benchmark representation

The primary benchmark starts from amg_dc_corrected_aligned. Each complete trial is transformed with a zero-phase Butterworth 5–100-Hz band-pass filter using forward–backward second-order-section filtering before window extraction. The exact implementation is:

sos = scipy.signal.butter(
    N=4, Wn=[5.0, 100.0], btype="bandpass", fs=1600.0, output="sos"
)
amg_5_100 = scipy.signal.sosfiltfilt(
    sos, complete_trial, axis=0, padtype="odd", padlen=27
)

SciPy N=4 produces four SOS sections after the band-pass transformation. The operation is offline and zero phase; it must not be interpreted as an online estimator. The 5–100-Hz representation was adopted as the primary benchmark input to attenuate low-frequency gross-motion components. It is not described as pure MMG and does not establish removal of all motion artifacts.

Each AMG input window contains 480 samples × 48 channels over 300 ms. Absolute time anchors are spaced by 100 ms. Random Forest inputs contain eight statistics for each AMG channel (384 features). The matched EMG path uses a 300-ms physical-time window containing 300 samples × four channels and four Hudgins features per channel (MAV, WL, ZC, and SSC), giving 16 features. AMG+EMG fusion therefore has 400 features.

Target alignments and window inventory

Before target validity is applied, the benchmark contains 264,081 base window identities. AMG and EMG cover the same identities with no modality-specific window loss.

Target alignment	All valid	Within train	Within validation	Within test	Cross train	Cross validation	Cross test
Window-end estimation	263,999	132,123	65,957	65,919	176,229	26,083	61,687
Centered-window offline estimation	263,994	132,119	65,955	65,920	176,225	26,083	61,686
100-ms-ahead target alignment	263,833	132,036	65,920	65,877	176,129	26,062	61,642

The count differences arise only from the validity of the requested native MoCap target frame and its four joint-angle values. Machine-readable counts and public-ID window manifests are under source_data/window_indices/ and source_data/benchmark/.

Fixed reference benchmark

The primary table uses all valid cross-subject window-end test windows. MAE, RMSE, and Pearson r are computed separately for each of the four outputs and then averaged arithmetically across outputs.

Model	MAE (degrees)	RMSE (degrees)	Pearson r
Random Forest	9.591	13.639	0.634
TCN	8.840	12.515	0.672
LSTMPlus	9.535	13.589	0.630
InceptionTimePlus	9.426	12.932	0.658

All four rows use fixed seed-42 reference predictions. Prediction-derived source values are published in source_data/benchmark/primary_cross_subject_window_end.csv.

Frequency representations

The frequency analysis contains four conditions generated independently from the same complete-trial, DC-corrected aligned base signal:

• DC-corrected full-band representation, with no additional frequency-selective filter;
• 0–20-Hz low-pass representation;
• 5–100-Hz filtered AMG representation, the primary representation;
• 100–760-Hz exploratory band-pass representation.

The conditions share the same 61,687 cross-subject window-end test identities. The 100–760-Hz condition is an analysis of a digital representation and does not establish the effective sensor bandwidth.

Auditory cue records

Cue records are published for 1,726 of the 1,918 retained trials; 192 retained trials have no cue record. The absent records are exactly the 64 trials for each of user05, user14, and user20. Every available record contains three cue onsets. Two excluded user23 trials had source cue records, but their relative cue-onset fields are blank because no released trial origin exists for an excluded trial.

The cue timestamps are protocol annotations relative to the shared trial origin. They were not used for alignment, target construction, windowing, splitting, model fitting, or metric calculation. See AUDITORY_CUE_METADATA.md.

Repository structure

data/trials/                  1,918 released NPZ trial files
metadata/                     subject, condition, channel, trial, split, and cue metadata
code/                         version-2.0.1 benchmark utilities and compatibility aliases
models/                       64 fitted 5–100-Hz v3 model artifacts and supporting records
source_data/model_registry/   70-condition registry: 64 fits plus six reuse mappings
source_data/window_indices/   public-ID window and target manifests
source_data/benchmark/        benchmark, frequency, and modality inventory tables
source_data/figures/          fixed-reference source tables for Figs. 3–5
manifests/                    hashes and release contracts

See RELEASE_NOTES_v2.0.2.md for the current Viewer patch and RELEASE_NOTES_v2.0.1.md for the naming-compatibility patch. The submitted article and Supplementary Information are not dataset files and are not included in this repository.

Citation, licence, and access

The dataset is publicly accessible at the repository URL above and is persistently identified by DOI 10.57967/hf/9950. Preferred citation metadata are provided in CITATION.cff.

No dataset licence, model-artifact licence, or software/code licence has been assigned in this repository version. Public accessibility does not itself define reuse permission. See LICENCE_STATUS.md.