Statistical and Machine Learning Models for Classification of Human Wear and Delivery Days in Accelerometry Data

Area of Science:

• Biomedical Engineering
• Data Science
• Wearable Technology

Background:

• Accelerometers are vital tools in biomedical research for objective physical activity assessment.
• Analysis of accelerometry data is challenged by large dataset sizes and unwanted data from device delivery.
• Manual review of delivery data is time-consuming and labor-intensive.

Purpose of the Study:

• To develop automated models for classifying accelerometry data.
• To distinguish between human wear activity and delivery process data.
• To streamline the cleaning of accelerometry datasets corrupted by delivery artifacts.

Main Methods:

• Developed statistical and machine learning models for supervised classification.
• Utilized a large dataset labeled for human activity and delivery.
• Assessed model performance using Monte Carlo cross-validation.

Main Results:

• A hybrid convolutional recurrent neural network achieved the highest performance (F1 score: 0.960).
• Simpler models like logistic regression (F1: 0.951) and random forest (F1: 0.957) also demonstrated high accuracy.
• Models effectively automate the identification and removal of delivery-related data.

Conclusions:

• Automated classification models significantly improve the efficiency of processing accelerometry data.
• Both complex deep learning and simpler machine learning models are effective for this task.
• The developed models and techniques are available in the R package 'Physical Activity' for broader use.