Validation of an Activity Type Recognition Model Classifying Daily Physical Behavior in Older Adults: The HAR70+...

Area of Science:

• Gerontology
• Biomedical Engineering
• Machine Learning

Background:

• Activity monitoring using machine learning (ML) offers insights into older adults' physical behavior.
• Existing ML models trained on young adults may not accurately reflect older adults' activities.
• Older adults exhibit diverse physical functions and may use walking aids, impacting activity recognition.

Purpose of the Study:

• To evaluate an existing ML model (HARTH) for classifying older adults' physical behavior.
• To compare HARTH with a new ML model (HAR70+) trained on older adult data.
• To assess ML model performance in older adults with and without walking aids.

Main Methods:

• Eighteen older adults (70-95 years) participated in a free-living protocol with chest cameras and accelerometers.
• Video analysis provided ground truth for activity recognition (walking, standing, sitting, lying).
• Two ML models (HARTH and HAR70+) were trained and evaluated on accelerometer data.

Main Results:

• Both HARTH (91%) and HAR70+ (94%) achieved high overall accuracy.
• Model performance decreased for participants using walking aids.
• The HAR70+ model improved accuracy from 87% to 93% for walking aid users.

Conclusions:

• The HAR70+ model demonstrates superior accuracy in classifying daily physical behavior in older adults.
• Accurate activity recognition is crucial for future research on older adult physical behavior.
• This validated model supports better understanding and monitoring of physical activity in the elderly.