Wrist-Based Fall Detection: Towards Generalization across Datasets.

Area of Science:

• Gerontology
• Biomedical Engineering
• Machine Learning

Background:

• Falls are a major cause of reduced independence and significant healthcare costs in older adults.
• Existing fall detection systems often use limited datasets, leading to poor generalization and high false positive rates.
• There is a need for more robust and comprehensive datasets for training reliable fall detection models.

Purpose of the Study:

• To develop a novel, larger dataset for fall detection by integrating data from three distinct sources.
• To establish a standardized method for incorporating future datasets, enhancing scalability.
• To evaluate the effectiveness of the combined dataset using cost-sensitive machine learning techniques.

Main Methods:

• A new dataset was created by merging three existing datasets, resulting in over 1300 fall samples and 28,000 non-fall samples.
• Temporal and frequency features were extracted from accelerometer and gyroscope data using a 2-second sliding window with 50% overlap.
• Cost-sensitive machine learning models were employed to address class imbalance inherent in fall detection data.

Main Results:

• The combined dataset demonstrated superior generalization capabilities compared to individual datasets.
• The developed model achieved high performance metrics: 90.57% recall, 96.91% specificity, and 98.85% Area Under the ROC Curve (AUC-ROC).
• The model effectively distinguished between daily activities and fall events.

Conclusions:

• The proposed integrated dataset significantly enhances the performance and reliability of fall detection systems.
• This approach addresses the limitations of small, single-source datasets, paving the way for more accurate real-world applications.
• The developed dataset and methodology offer a scalable solution for improving fall prevention strategies in aging populations.