Smartwatch IMU Dataset for Upper-Limb Movement Analysis: Raw, Processed, and Feature Matrices

This dataset contains dual-wrist Apple Watch (Series 6) inertial recordings for three canonical upper-limb movements—elbow flexion/extension (el-exfl), shoulder flexion/extension (sh-exfl), and wrist pronation/supination (wr-prsu)—collected from three adults (IDs: u01, u04, u05; right-hand dominant; one with mild left-arm weakness). Signals were sampled at 20 Hz via the HemiPhysioData watchOS app using CoreMotion and saved as synchronized CSV time series.

The repository provides:
- Raw data (raw/): user acceleration, rotation rate, gravity, Euler orientation (roll, pitch, yaw), and quaternions; per-session CSVs with metadata (UID, Wrist, Side, MoveType, SessionID).
- Processed data (processed/): standardized/smoothed streams and overlapping sliding-window segments (2.56 s and 1.0 s, 50% overlap).
- Features (features/): time- and frequency-domain feature matrices for Set_1 (432), Set_2 (576), Set_3 (720) features per window, plus the selected 46-feature subset.
- Labels (labels/): segment-level metadata table.
- Docs (docs/): data dictionary (columns.md), provenance (provenance.json), and README.

Use cases include benchmarking human-activity recognition (HAR) for upper-limb tasks, ablation studies on sensor modality and window length, and reproducible pipelines for feature extraction and feature selection. Data are released under CC BY 4.0. Please cite this dataset’s DOI and the associated article:
Benachour, Y., Rehman, M., & Flitti, F. Towards improved human arm movement analysis: Advanced feature engineering and model optimization, Engineering Applications of Artificial Intelligence, 156:111194, 2025.

Keywords: smartwatch; IMU; upper-limb; rehabilitation; human activity recognition; feature engineering; sliding-window; PCA; t-SNE; Apple Watch.