Physical Activity Monitoring in Older Adults (80+) in Simulated Free-Living Conditions (Synthetic Data)

This is the synthetic data generated from the real data, available at DOI: 10.5281/zenodo.18222937 The synthetic dataset was generated based on the real measurements using a signal-processing-based method. The proposed synthetic data generation scheme has been published in a conference paper submitted to the Nordic Conference on Digital Health and Wireless Solutions 2026, whose doi: https://doi.org/10.1007/978-3-032-28819-6_9 The summarised methodology is presented below: Synthetic data were generated by extracting generalized gait representations that capture common characteristics across individuals while minimizing participant-specific biases. First, raw acceleration signals were visually inspected for anomalies. For walking, large baseline drifts along the APDM z-axis were observed in a small number of participants. As no kinematically plausible explanation could be identified, these drifts were removed using a noncausal fourth-order high-pass filter with a cut-off frequency of 0.5~Hz. Second, generalized activity representations were generated using dynamic time warping barycentre averaging (DBA). DBA is a DTW-based averaging technique that aligns time series with varying temporal dynamics by nonlinear stretching or compression, while enforcing alignment of the first and last samples. After alignment, the barycentre of the aligned signals is computed, yielding a representative time series that captures common temporal patterns despite inter-individual variability. Synthetic windows were generated for walking, standing, sitting, supine lying, left and right lateral recumbent lying, sit-to-stand, stand-to-sit, sit-to-lie, and lie-to-sit transfers. To ensure a fair comparison, the number of 2-second windows per activity was matched between the real and synthetic datasets, preserving identical class imbalance conditions.