Application of Convolutional Neural Network Algorithms for Advancing Sedentary and Activity Bout Classification.

Area of Science:

• Wearable sensor technology
• Human activity recognition
• Machine learning for health

Background:

• Classifying physical behavior from accelerometers is challenging in real-world settings.
• Deep learning, specifically CNNs, may better represent complex, free-living data without engineered features.
• Previous machine learning models faced limitations with observational data.

Purpose of the Study:

• To develop a modeling pipeline for evaluating a CNN on free-living accelerometer data.
• To compare CNN performance against random forest and logistic regression.
• To assess the impact of feature engineering on model accuracy.

Main Methods:

• 28 women wore hip-mounted accelerometers and thigh-mounted activity monitors for 7 days.
• Data were used to train and evaluate logistic regression, random forest, and CNN models.
• Models classified bouts of sitting, standing, and stepping.

Main Results:

• The CNN model achieved the highest accuracy (84%) for classifying activity bouts.
• CNN performance surpassed logistic regression (56%) and random forest (76%).
• Feature engineering did not improve CNN performance.

Conclusions:

• CNNs demonstrate superior performance in classifying free-living physical activity compared to traditional methods.
• Deep learning models show promise for handling complex, real-world behavioral data.
• CNNs offer potential for improved transferability to diverse populations.