NT-FDS-A Noise Tolerant Fall Detection System Using Deep Learning on Wearable Devices

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• Unintentional falls in the elderly are a significant public health concern with severe consequences.
• Fall Detection Systems (FDS) are crucial for timely medical intervention but are hampered by data integrity issues, particularly missing values.
• Missing data in FDS can arise from sensor unreliability, interference, or data transmission problems, compromising system effectiveness.

Purpose of the Study:

• To develop and evaluate a noise-tolerant Fall Detection System (FDS) capable of performing reliably in the presence of missing data.
• To leverage Deep Learning (DL), specifically Recurrent Neural Networks (RNNs) with Bidirectional Long Short-Term Memory (BiLSTM), for enhanced fall detection using wearable sensors.
• To assess the proposed FDS against existing state-of-the-art methods on benchmark datasets.

Main Methods:

• Implementation of a Fall Detection System (FDS) utilizing a Bidirectional Long Short-Term Memory (BiLSTM) neural network architecture.
• Training and evaluation of the FDS on two publicly available datasets: SisFall and UP-Fall Detection.
• Comparative analysis of the proposed BiLSTM-based FDS performance against current state-of-the-art fall detection techniques.

Main Results:

• The proposed BiLSTM-based FDS achieved high performance metrics: accuracy of 97.21% (SisFall) and 97.41% (UP-Fall Detection).
• Exceptional sensitivity was recorded at 96.97% (SisFall) and 99.77% (UP-Fall Detection), indicating robust fall identification.
• Specificity reached 93.18% (SisFall) and 91.45% (UP-Fall Detection), demonstrating effective discrimination between falls and normal activities.
• The system outperformed existing state-of-the-art methods on both benchmark datasets, particularly in handling missing data.

Conclusions:

• The Bidirectional Long Short-Term Memory (BiLSTM) model is highly effective for wearable fall detection systems, especially when dealing with missing data.
• The inherent ability of BiLSTM to capture long-term dependencies from both past and future data makes it suitable for robust fall detection.
• The developed FDS offers a promising solution for improving the reliability and effectiveness of fall detection in elderly populations, even under challenging data conditions.