Performance Analysis of Data Augmentation Approaches for Improving Wrist-Based Fall Detection System.

Area of Science:

• Gerontology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Societal aging increases fall risks in the elderly, leading to severe physical, psychological, and financial consequences.
• Effective fall detection systems are crucial for timely alerts and mitigating fall-related harm.
• Wrist-based systems offer convenience but face performance challenges due to complex hand motion modeling and data limitations.

Purpose of the Study:

• To investigate and compare various data augmentation techniques for improving deep learning-based wrist-worn fall detection systems.
• To address the common issues of class imbalance and data scarcity in fall detection datasets.
• To identify the most effective data augmentation method for enhancing the performance of elderly fall detection.

Main Methods:

• Analysis of multiple data augmentation methodologies applied to wrist-based sensor data for fall detection.
• Implementation of deep learning models trained on augmented datasets.
• Evaluation of system performance using metrics such as the F1 score, particularly under conditions of limited training data.

Main Results:

• The conditional diffusion model demonstrated superior performance as a data augmentation technique.
• The F1 score improved by 6.58% when the model was trained using only 25% of the original data.
• Generated synthetic data maintained high quality, effectively supplementing the limited real-world data.

Conclusions:

• The conditional diffusion model is a highly effective approach for data augmentation in wrist-based fall detection systems.
• This method significantly enhances fall detection accuracy, especially when dealing with scarce data, making it ideal for elderly fall monitoring.
• High-quality synthetic data generation can overcome data limitations, improving the reliability of deep learning models for fall detection.