Event-Centered Data Segmentation in Accelerometer-Based Fall Detection Algorithms.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Gerontology

Background:

• Automatic fall detection systems are crucial for elderly assistance.
• Accelerometer-based systems offer portability and low cost but struggle to differentiate falls from Activities of Daily Living (ADL).
• Improved accuracy is needed for widespread adoption of fall detection technology.

Purpose of the Study:

• To investigate the impact of data segmentation window configurations on fall detection accuracy.
• To identify the optimal window settings for accelerometer-based fall detection systems.
• To enhance the reliability of distinguishing falls from ADL.

Main Methods:

• Developed a testing environment using a Support Vector Machine (SVM) classifier.
• Evaluated various numbers and durations of data segmentation windows.
• Employed an event-centered approach, defining windows relative to potential fall events.
• Utilized three publicly available datasets containing fall and ADL data.

Main Results:

• A three-sequential-window configuration (pre-impact, impact, post-impact) yielded the highest detection accuracy across all datasets.
• Optimal performance was achieved with a 0.5s or 1s impact window.
• Combined with 3.5s or 3.75s pre- and post-impact windows, this configuration maximized accuracy.

Conclusions:

• Data segmentation is a critical factor in accelerometer-based fall detection.
• The proposed three-window segmentation strategy significantly enhances fall detection accuracy.
• This research provides optimal parameters for improving real-world fall detection systems.