Application of machine learning to predict transport modes from GPS, accelerometer, and heart rate data

Area of Science:

• Human-computer interaction
• Transportation science
• Machine learning for health

Background:

• Measuring active transport is challenging and prone to errors.
• Sensor data, including Global Positioning System (GPS) and accelerometer, are increasingly used for transport mode prediction.
• Heart rate data's utility in transport mode prediction has been largely unexplored.

Purpose of the Study:

• To predict transport modes using a combination of Global Positioning System (GPS), accelerometer, and heart rate data.
• To investigate the impact of methodological choices, such as data splitting and post-processing, on prediction accuracy.
• To evaluate the contribution of heart rate data to transport mode prediction.

Main Methods:

• The RECORD MultiSensor study collected sensor data (GPS, accelerometer, heart rate) from 126 participants over seven days.
• Random Forests (RF) models were trained and tested at the participant level to predict transport modes per minute.
• Various window sizes for moving average post-processing were evaluated to homogenize predictions.

Main Results:

• Minute-level prediction accuracy for distinguishing between trips and locations reached 90%.
• Overall transport mode prediction rates varied, from 65% for public transport to 95% for biking.
• Heart rate data provided only a marginal improvement in prediction accuracy, primarily for biking.
• Participant-level data splitting, rather than minute-level, was crucial to avoid biased prediction rates.

Conclusions:

• Careful definition of training and test sets in Random Forests (RF) models is essential for reliable transport mode prediction.
• Post-processing using optimized moving average windows can enhance prediction accuracy.
• Heart rate data offers limited added value for predicting most transport modes.