A Study on the Application of Convolutional Neural Networks to Fall Detection Evaluated with Multiple Public Datasets

Area of Science:

• Biomedical Engineering
• Gerontology
• Artificial Intelligence

Background:

• Falls significantly impact older adults' health, independence, and healthcare costs.
• Wearable fall detection systems (FDSs) are crucial for mitigating these consequences.
• Effective algorithms are needed to distinguish falls from Activities of Daily Life (ADLs).

Purpose of the Study:

• To present and evaluate a convolutional deep neural network (CNN) for fall pattern identification.
• To assess the CNN's performance using measurements from a tri-axial accelerometer.
• To evaluate the model's generalizability across diverse public fall detection datasets.

Main Methods:

• A convolutional deep neural network (CNN) architecture was developed and applied.
• The CNN was trained and tested on data from a transportable tri-axial accelerometer.
• Evaluation involved multiple public data repositories with ADLs and mimicked falls from various volunteers.

Main Results:

• The CNN achieved high accuracy when hyper-parameterized for specific datasets.
• Global evaluation across different repositories revealed challenges in model generalizability.
• Optimized network architectures for one dataset did not consistently perform well on others.

Conclusions:

• Convolutional deep neural networks show promise for fall detection using accelerometer data.
• Hyper-parameter optimization is critical for achieving high performance on specific datasets.
• Significant challenges remain in developing fall detection algorithms that generalize across diverse testing environments and populations.