TOWARDS MUSCULOSKELETAL SIMULATION-AWARE FALL INJURY MITIGATION: TRANSFER LEARNING WITH DEEP CNN FOR FALL DETECTION

Area of Science:

• Geriatric Medicine
• Biomedical Engineering
• Machine Learning

Background:

• Fall detection is crucial for mitigating injuries in geriatric populations.
• Existing methods struggle with limited labeled sensor data.
• Deep convolutional neural networks (DCNNs) show promise in image recognition.

Purpose of the Study:

• To develop an effective fall detection method for geriatric subjects.
• To leverage transfer learning for improved accuracy with scarce data.
• To integrate machine learning with musculoskeletal modeling for fall injury mitigation.

Main Methods:

• Utilized a pre-trained kinematics-based machine learning approach.
• Converted accelerometry data into images using time-frequency analysis (scalograms) via continuous wavelet transform.
• Applied data augmentation to scalogram images to enhance model performance.
• Employed transfer learning from existing large-scale annotated accelerometry datasets.

Main Results:

• Transfer learning demonstrated superior performance compared to existing methods.
• The method showed improved accuracy, particularly with limited labeled training data.
• Experimental results were validated on the publicly available URFD dataset.

Conclusions:

• Transfer learning is a viable and effective strategy for fall detection using accelerometry data.
• The proposed image-based approach enhances fall detection accuracy in data-scarce scenarios.
• This work contributes to the development of fall injury mitigation systems for the elderly.