Identification of Travel Behaviour from Objective Physical Activity Data. (June 2017)

• Record Type: Journal Article
• Authors: Procter, Duncan; Nightingale, Claire; Rudnicka, Alicja; Ram, Bina; Ellaway, Anne; Cooper, Ashley; Page, Angie; Lewis, Daniel; Cummins, Steven; Giles-Corti, Billie; Whincup, Peter; Cook, Derek; Owen, Christopher
• Is Part Of: Journal of transport & health. Volume 5(2017)Supplement
• Page Start: S39
• Publication Date: 2017-06
• Subjects: Transportation -- Health aspects -- Periodicals; Transportation -- Periodicals; Public Health -- Periodicals; Noise, Transportation -- Periodicals; Air Pollutants -- Periodicals; 388
• DOI: 10.1016/j.jth.2017.05.326
• Languages: English
• ISSNs: 2214-1405

Abstract

Abstract : Background: Current UK health guidelines suggest a minimum of 150 minutes of moderate intensity physical activity per week to decrease the risk of several non-communicable diseases. Active travel, i.e. travel powered by physical activity such as walking or cycling, represents an opportunity to increase people's overall level of activity without dramatic changes to other parts of their lifestyle. For measurements of activity to be reliable they must be objective. Physical activity is frequently assessed objectively using accelerometers, with their location determined by GPS devices. However, if travel behaviour is to be properly understood, this accelerometer and GPS data needs to be separated into different travel modes. The ENABLE study (Examining Neighbourhood Activities in Built Living Environments) is a longitudinal natural experiment based in London, UK, which is assessing the effects of moving to a physical activity-supportive environment on participants' travel behaviour, and which offers an excellent opportunity to develop automated methods to identify and quantify travel behaviour. Methods: At baseline, 1089 adults had their physical activity assessed using accelerometers and GPS devices. We present a method for the identification of five travel modes from this baseline data (Walking, Cycling, Vehicles, Trains, Not travelling), using accelerometer counts, GPS-derived speed data and freely available GIS data. We use a supervised-machine learning algorithm based on Bayesian networks. We present the creation of training data from survey questionnaires, data processing, network training and prediction and post-processing of output. In post-processing we also identify points at which underground train journeys may have taken place, removing a potential bias of GPS-based objective measurement of travel behaviour. Results: Our trained Bayesian network predicts five travel modes with 88.4% correct prediction across all travel modes (lowest cycling 75.0%, highest walking 94.3%) After application of post-processing rules, this increases to 90.1% overall correct prediction (lowest cycling 77.2%, highest walking 95.0%). Underground train journeys, which give no GPS signal, are identified using simple rules. A total of 3154 potential journeys are identified, of which 2343 were correct after manual validation. Conclusions: Our identification method successfully distinguishes between five different travel modes with high accuracy, using easily attainable objective data and is implemented in the free statistical software R. Our method removes the need for the collection of time consuming training data, and allows analysis of differences between active and passive travel methods.