Accelerometer-Based Machine Learning Categorization of Body Position in Adult Populations

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Human Movement Science

Background:

• Current sensor technology may lack data granularity for detailed movement analysis.
• Clinicians question the validity of movement analysis for specialized populations like older adults.
• Accurate movement classification is crucial for sensor-based clinical decision-making in healthcare settings.

Purpose of the Study:

• To evaluate classification algorithms using accelerometer data for posture movement detection.
• To compare the performance of single versus dual sensor setups.
• To assess algorithm accuracy across different age groups (adults <55 and older adults ≥55).

Main Methods:

• Developed custom software and classification algorithms to identify five posture movements: laying, reclining, sitting, standing, and walking.
• Collected accelerometer data from healthy adults and older adults.
• Tested algorithm performance with varying sensor configurations and populations.

Main Results:

• Achieved high classification accuracy: 93.2% for adults under 55 and 95% for older adults over 55.
• Demonstrated that sensor body position is critical for algorithm training and application.
• Indicated potential challenges in applying algorithms trained on one population to another.

Conclusions:

• The developed custom algorithms show high accuracy for classifying posture movements in healthy adults and older adults.
• Findings suggest the need for careful consideration of sensor placement and population-specific algorithm training.
• This approach can facilitate future research on movement classification in hospitalized older adults.