A cross-dataset deep learning-based classifier for people fall detection and identification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Engineering

Background:

• Falls pose a significant risk to elderly individuals living alone, necessitating automated detection and assistance systems.
• Current systems may require dataset-specific fine-tuning, limiting their real-world applicability.
• Prompt identification of individuals after a fall is crucial for emergency response.

Purpose of the Study:

• To develop a novel deep learning approach for simultaneous fall detection and subject identification.
• To create a dataset-independent model that requires no fine-tuning for new datasets or subjects.
• To enhance the safety and rapid response for vulnerable populations prone to falls.

Main Methods:

• A multi-task learning deep learning model was developed, utilizing raw inertial data as input.
• The model simultaneously addresses fall detection and subject identification tasks.
• The approach learns optimal data representations automatically, avoiding pre-processed feature constraints.

Main Results:

• The cross-dataset classifier achieved over 98% accuracy in fall detection across four diverse datasets.
• False positive rates were minimal, averaging less than 1.6%, setting a new state-of-the-art.
• The model demonstrated an average accuracy of 79.6% for identifying individuals.

Conclusions:

• A single deep learning model can effectively perform both fall detection and subject identification in real-time.
• The proposed method is robust across different datasets and conditions without requiring retraining.
• This approach simplifies implementation and improves the reliability of automated fall monitoring systems.