Refining Time-Activity Classification of Human Subjects Using the Global Positioning System.

Area of Science:

• Environmental Health
• Exposure Science
• Data Science

Background:

• Accurate personal exposure estimation to air pollution relies on detailed spatial location data.
• Global Position System (GPS) is crucial for tracking personal movement and activities.
• Existing time-activity classification models are often region-specific, limiting broad application.

Purpose of the Study:

• To develop an adaptive, automated model for classifying time-activity patterns using GPS data.
• To enhance the applicability of GPS-based individual activity tracking in air pollution studies.
• To assess the contribution of accelerometer and spatial data to classification accuracy.

Main Methods:

• Collected 7-day time-activity data from 13 adults in Southern California using GPS loggers and accelerometers.
• Developed a random forests model to classify activities into indoor, outdoor-static, outdoor-walking, and in-vehicle travel.
• Performed sensitivity analysis and cross-validation (leave-one-fold-out, leave-one-subject-out) to evaluate model performance.

Main Results:

• The random forests model achieved high accuracy (99.71% overall) using GPS, roadway, and tax parcel data.
• Maximum speed and distance to highways were key predictive variables.
• Excluding accelerometer data had minimal impact on overall accuracy, but leave-one-subject-out validation revealed reduced sensitivity for certain outdoor activities.

Conclusions:

• The random forests model accurately classifies major time-activity categories using GPS and spatial data.
• The model shows promise for air pollution exposure studies.
• Generalizability to diverse populations requires careful consideration due to the study's small sample size.