Using Deep Learning Predictions of Smokers’ Behaviour to Develop a Smart Smoking-Cessation App

Abo-Tabik, M. A. (2021) Using Deep Learning Predictions of Smokers’ Behaviour to Develop a Smart Smoking-Cessation App. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (7MB) | Preview

Abstract

The number of new smoking-cessation apps had increased in recent years. Although these offer accessible and low-cost support to smokers, they often lack scientific understanding of nicotine addiction, and rely on smokers’ self-reporting their cravings / environmental factors; a method widely acknowledged to be unreliable. This PhD presents two novel deep-learning models for automatic smoking events prediction. Both models combine machine-learning with Control Theory Model of Smoking (CTMoS), to enable the prediction of smoking events based on both internal (nicotine level) and external (e.g. location) factors. This offers a way to overcome limitations of previous apps. The first model, combined CTMoS with a 1D Convolutional Neural Network, using raw accelerometer and GPS coordinates as input. Result indicated good prediction of internal craving factors (e.g. nicotine level and craving); but smoking events prediction required improvement, as the f1-score were 0.06, 0.14, 0.24, and 0.4 for predicting a smoking event 5, 15, 30, and 60 -min (respectively) prior to its occurrence. The second model combined 1D Convolutional Neural Network with the Bidirectional Long Short-Term Memory method, to create a deep learning model with Genetic Algorithm for hyperparameter selection. The model used the same 3- accelerometer values as input, but the 3-GPS coordinates were replaced with coded location data (five most smoked locations). These changes improved smoking events prediction with average f1-score of 0.32, 0.59, 0.71, and 0.8 for predicting a smoking event 5, 15, 30, and 60 -min (respectively) prior to its occurrence. This PhD achieved its three goals: minimize user input (by using data collected from phone sensors); improve scientific understanding of factors that influence smokers’ behaviour (by evaluating the relative contribution of different factors), and developing a state-of-the-art model that enables the automatic prediction of smoking events. As such, outcomes of this PhD lay the foundation for future development of smart and personalized apps that can provide real-time personalized support for smokers.