Ordinal Statistical Models of Physical Activity Levels from Accelerometer Data.

Area of Science:

• Biomedical Engineering
• Data Science
• Wearable Technology

Background:

• Advancements in accelerometer technology enable sophisticated physical activity assessment.
• Predictive modeling requires robust methods to interpret complex sensor data.

Purpose of the Study:

• To implement and evaluate ordinal random forest and partial proportional odds models for physical activity assessment using accelerometer data.
• To compare the performance of these models against traditional random forest approaches.

Main Methods:

• Utilized accelerometer data from 28 adults performing daily activities across two visits (training and testing).
• Employed an independent sample, cross-validation approach.
• Applied ordinal random forest and partial proportional odds models accounting for response ordinality.

Main Results:

• Ordinal random forest demonstrated comparable accuracy and superior linearly weighted kappa values to standard random forest.
• The ankle sensor yielded the highest accuracy (33.3%) for a four-activity level classification.
• Error rates for a two-activity level classification were lowest on the ankle (15.5%).

Conclusions:

• Ordinal random forest and partial proportional odds models are effective for assessing physical activity from accelerometer data.
• These models offer objective activity level assessment without direct observation, applicable to exercise studies.
• Findings suggest potential for theoretical advancements in current physical activity modeling techniques.