Analysis of Public Datasets for Wearable Fall Detection Systems

Area of Science:

• Biomedical Engineering
• Signal Processing
• Human-Computer Interaction

Background:

• Wearable Fall Detection Systems (FDSs) are crucial due to the proliferation of smart devices.
• Evaluating FDS effectiveness requires diverse datasets of falls and Activities of Daily Living (ADLs) from inertial sensors.
• Publicly accessible databases are fundamental for the systematic assessment of fall detection algorithms.

Purpose of the Study:

• To review and critically appraise twelve existing public datasets for evaluating wearable FDSs.
• To identify limitations and heterogeneity in current fall and ADL datasets.
• To propose improvements for future dataset development and FDS evaluation.

Main Methods:

• Comprehensive review and analysis of twelve public datasets for FDS research.
• Evaluation of datasets based on testbed generation factors, sensor characteristics, and data typologies.
• Statistical analysis of sensor range impact on data reliability.

Main Results:

• Significant heterogeneity exists across datasets regarding sample size, fall/ADL types, subject characteristics, and sensor configurations.
• A lack of standardized experimental benchmarking procedures hinders direct comparison of FDS algorithms.
• Sensor range significantly impacts the reliability of collected inertial data.

Conclusions:

• Standardized methodologies for generating and reporting FDS datasets are urgently needed.
• Categorizing ADLs by movement intensity is essential for accurately evaluating fall detection algorithms' ability to discriminate falls.
• Future research should focus on creating more robust, standardized, and diverse datasets to advance wearable FDS technology.