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

Abstract

Background: There has been an increased focus on active transport, but the measurement of active transport is still difficult and error-prone. Sensor data have been used to predict active transport. While heart rate data have very rarely been considered before, this study used random forests (RF) to predict transport modes using Global Positioning System (GPS), accelerometer, and heart rate data and paid attention to methodological issues related to the prediction strategy and post-processing.

Methods: The RECORD MultiSensor study collected GPS, accelerometer, and heart rate data over seven days from 126 participants living in the Ile-de-France region. RF models were built to predict transport modes for every minute (ground truth information on modes is from a GPS-based mobility survey), splitting observations between a Training dataset and a Test dataset at the participant level instead at the minute level. Moreover, several window sizes were tested for the post-processing moving average of the predicted transport mode.

Results: The minute-level prediction rate of being on trips vs. at a visited location was 90%. Final prediction rates of transport modes ranged from 65% for public transport to 95% for biking. Using minute-level observations from the same participants in the Training and Test sets (as RF spontaneously does) upwardly biases prediction rates. The inclusion of heart rate data improved prediction rates only for biking. A 3 to 5-min bandwidth moving average was optimum for a posteriori homogenization.

Conclusion: Heart rate only very slightly contributed to better predictions for specific transport modes. Moreover, our study shows that Training and Test sets must be carefully defined in RF models and that post-processing with carefully chosen moving average windows can improve predictions.