Using GPS, GIS, and Accelerometer Data to Predict Transportation Modes

作者：Brondeel Ruben^*; Pannier Bruno; Chaix Basile

来源：Medicine and Science in Sports and Exercise, 2015, 47(12): 2669-2675.

DOI：10.1249/MSS.0000000000000704

• PHYSICAL ACTIVITY
• ACTIVE TRANSPORT
• PASSIVE DATA COLLECTION
• MACHINE LEARNING
• RECORD COHORT STUDY
• FRANCE

Introduction: Active transportation is a substantial source of physical activity, which has a positive influence on many health outcomes. A survey of transportation modes for each trip is challenging, time-consuming, and requires substantial financial investments. This study proposes a passive collection method and the prediction of modes at the trip level using random forests. Methods: The RECORD GPS study collected real-life trip data from 236 participants over 7 d, including the transportation mode, global positioning system, geographical information systems, and accelerometer data. A prediction model of transportation modes was constructed using the random forests method. Finally, we investigated the performance of models on the basis of a limited number of participants/trips to predict transportation modes for a large number of trips. Results: The full model had a correct prediction rate of 90%. A simpler model of global positioning system explanatory variables combined with geographical information systems variables performed nearly as well. Relatively good predictions could be made using a model based on the 991 trips of the first 30 participants. Conclusions: This study uses real-life data from a large sample set to test a method for predicting transportation modes at the trip level, thereby providing a useful complement to time unit-level prediction methods. By enabling predictions on the basis of a limited number of observations, this method may decrease the workload for participants/researchers and provide relevant trip-level data to investigate relations between transportation and health.

• 出版日期2015-12