Online Fall Detection Using Wrist Devices.

Area of Science:

• Gerontology
• Biomedical Engineering
• Machine Learning

Background:

• Falls are a major health concern for the elderly, leading to fear of falling, reduced mobility, and decreased quality of life.
• Existing wrist-based fall detection datasets lack realistic scenarios and data from elderly individuals.
• Addressing these limitations is crucial for developing effective fall prevention and monitoring technologies.

Purpose of the Study:

• To introduce a novel wrist-based dataset specifically designed for fall detection in the elderly.
• To develop and evaluate a low-power, accelerometer-only fall detection system using this dataset.
• To investigate the performance of machine learning algorithms with a computationally efficient feature set for real-time fall detection.

Main Methods:

• A new wrist-based dataset was created, focusing on realistic movements of elderly individuals.
• An accelerometer-only fall detector was developed, streaming data at 50 Hz.
• The FS-1 feature set (maximum, minimum, mean, variance) was used to train a 3-Nearest Neighbors (3NN) algorithm with Euclidean distance and a 9-second window size.
• A learning version of the fall detector was implemented to improve performance over time.

Main Results:

• The proposed system utilizes a computationally efficient FS-1 feature set and a 3NN algorithm.
• The fall detector achieved zero false positives and zero false negatives over a four-day testing period.
• The system was designed with battery and memory limitations in mind, suitable for wearable devices.

Conclusions:

• The developed wrist-based dataset and fall detection system show significant promise for enhancing the safety and independence of the elderly.
• The low-power, high-accuracy fall detection system addresses key limitations of previous approaches.
• Further development of learning-based fall detection systems can lead to improved real-time monitoring and intervention for falls in older adults.