Synthetic Multirate Time-Series Dataset for Health Monitoring and Wearable Fall Detection in Older Adults
Description
This dataset is a synthetic multirate wearable sensor time-series dataset designed for benchmarking health monitoring, fall detection, sensor-fault robustness, and machine learning tasks for older adults. It contains 120 simulated subjects aged 65–95 years, with balanced female and male profiles, and 960 four-second events (eight per subject). Subject-wise splits include 84 training, 18 validation, and 18 test subjects, supporting leakage-aware model development and evaluation. Seven wearable sensing configurations are included, covering waist-worn inertial sensors, wrist-based smartwatches, waist and pocket placements, lower-back inertial sensors, and smartphone-in-pocket sensing. Sampling rates range from 20 Hz to 200 Hz, with 80, 400, or 800 samples per event. The sample-level data contain 362,240 samples and 45 columns covering sensor measurements, labels, provenance, and event context. The dataset includes 12 daily activities and four fall types, together with hard-negative activities that can produce fall-like sensor patterns, including abrupt sitting, jumping, stumble recovery, and stepping into or out of a car. The release contains 124 falls and 337 hard-negative activities. Sensor-fault scenarios include inertial signal dropout and quantisation effects. Linked subject-, event-, and sample-level files are provided. Subject records contain demographic characteristics, sensor configuration, placement, sampling rate, baseline vital signs, and simulated health-condition indicators. Event records contain activity, health condition, sensor degradation, hard-negative status, fall information, simulated health-risk probability and labels, provenance, sample count, and peak acceleration. Sample-level Parquet data provide timestamps, event phase, three-axis acceleration, angular velocity, derived magnitudes, and physiological variables including heart rate, oxygen saturation, temperature, heart-rate variability, respiratory rate, and blood pressure. Derived resources include a 960 × 107 event-feature table and a 576 × 12 internal held-out prediction table. Documentation includes data dictionaries, sensor profiles, provenance, generation manifests, integrity audits, grouped-fold evaluation, subgroup analyses, domain-shift results, feature importance, and an external SisFall transfer stress test. Quality-control checks found no duplicate timestamp-sensor records, no duplicate event hashes, and no subject overlap between splits. Repeated six-axis vectors were associated with intentionally simulated sensor dropout or quantisation. Range checks passed, and event and prediction identifiers were complete. The dataset supports research in wearable sensing, time-series machine learning, fall detection, health-risk modelling, sensor-fault robustness, multirate signal processing, domain-shift analysis, and older-adult monitoring. Health-risk variables are simulation targets for benchmarking and should not be interpreted as clinical diagnoses.
Files
Steps to reproduce
1. Set up the environment: Run the supplied Google Colab notebook using Python 3.12.13 with NumPy, pandas, SciPy, scikit-learn, PyArrow, Matplotlib, seaborn, and gdown. Execute the 37 code cells sequentially. The release was generated using seed 2026. Exact package versions are provided in `environment_versions.csv`. 2. Initialize the generator: Use the documented generator configuration (generator version 3.1.0; release version 3.2.0). The default route uses frozen calibration values from a verified simulated parent snapshot; no patient observations are included. 3. Generate subjects: Create 120 simulated subjects (S0000–S0119), aged 65–95 years. Generate baseline physiological characteristics and simulated health-condition indicators. These probabilities are simulation settings, not population prevalence estimates. 4. Create subject-wise splits: Randomly divide subjects into 70% training, 15% validation, and 15% testing, resulting in 84, 18, and 18 subjects. Keep each subject and all events within the same split. 5. Generate events and sensor signals: Generate eight four-second events per subject, covering falls and daily activities. Use 20, 100, or 200 Hz according to the sensor profile. Generate four signal phases—pre-event, transition, impact/activity peak, and recovery—using Gaussian pulses, direction vectors, subject context, and scenario-specific ranges. Calculate acceleration and angular-velocity magnitudes. 6. Generate physiological data: Generate physiological measurements around each subject's baseline and modify them according to the simulated condition. Keep continuous health-risk probability, binary risk label, and observed health outcome as separate variables. 7. Apply sensor faults: Assign sensor quality conditions using documented probabilities for no degradation, noise, dropout, quantisation, and drift. For dropout, replace a continuous portion of the inertial signal with missing values. 8. Generate derived outputs: Calculate event-level statistical summaries and merge them with event and subject metadata to create `event_features.csv`. Generate sample-level Parquet data, metadata, predictions, documentation, provenance, and audit files. 9. Run quality checks: Verify signal ranges, missingness, subject consistency, split overlap, event duration, sampling intervals, duplicate records, repeated sensor vectors, event hashes, and boundary values. Interpret repeated vectors according to the sensor-degradation label. 10. Optional modelling and transfer tests: Reproduce the documented logistic-regression and random-forest examples using subject-separated data. For external transfer testing, use the verified SisFall archive, create four-second 800-sample windows with a 400-sample step, and apply the synthetic-data models without retraining. 11. Verify the release: Check the release using the provided manifest, inventory, provenance records, and SHA-256 checksums.