A comparison of public datasets for acceleration-based fall detection.

Area of Science:

• Gerontology
• Biomedical Engineering
• Computer Science

Background:

• Falls are a major cause of mortality in older adults, making rapid fall detection crucial.
• Acceleration-based fall detection using accelerometers or smartphones is an active research area.
• Existing studies report diverse fall detection rates, complicating comparisons.

Purpose of the Study:

• To compare publicly available fall detection datasets for the first time.
• To assess the influence of different datasets on fall detection algorithm performance.
• To evaluate the generalization capability of fall detection algorithms across datasets.

Main Methods:

• Utilized two distinct fall detection algorithms.
• Compared algorithm performance across multiple public fall detection datasets.
• Analyzed the impact of dataset characteristics (e.g., training samples, sampling frequency) on detection accuracy.

Main Results:

• Fall detection algorithm performance is significantly affected by the specific dataset used for validation.
• Generalization capability is poor; algorithms perform dramatically worse when tested on unseen datasets.
• The number of training samples influences performance, while sampling frequency and acceleration range have less impact.

Conclusions:

• Dataset choice critically impacts the reported performance of fall detection algorithms.
• Current fall detection models exhibit limited generalization across different datasets.
• Standardization or careful dataset selection is essential for reliable fall detection system development.