Title : Using mixed deep learning methods to model smokers behaviour from mobile phone location and movement data
Abstract:
In recent years a significant number of smoking-cessation apps have been developed to help smokers quit smoking. Although these apps offer easy access and low-cost support, most of them are designed without a scientific understanding of the complexity of nicotine addiction. They also mostly rely on the smoker to self-report their smoking craving and progress.
This project developed a model which can automatically predict smoking events based on data passively collected using a smartphone. The model combines deep learning methods with a Control Theory model of smoking to capture both internal and external smoking drivers.
The deep learning methods used in the design combine the 1D Convolutional Neural Network and the Bidirectional Long Short-Term Memory to take advantage of each method. It uses 1D Convolutional Neural Network to extract patterns and the Bidirectional Long Short-Term Memory to extract the sequential correlations in the input sequences. Additionally, to reach a near-optimal design, Genetic Algorithm are used for hyper-parameter selection; this model uses raw accelerometer values with codded GPS coordinates as an input.
Using the deep learning model and combining it with the Control Theory model of smoking, the model was able to predict smoking behaviour with an average smoking event prediction of 0.32, 0.59, 0.69, 0.76 for 5, 15, 30, and 60-min windows respectively, before the appearance of the smoking event.
The model opens the door to the possibility of automatic prediction of smoking behaviour, which can be used in the future to develop a smoking cessation app as it relies entirely only on data collected using smartphone devices.
