Whereas the majority of smokers will quit in any given year, the majority of quit attempts result in relapse. One reason interventions may fail is that they teach smokers strategies for coping with craving in response to environmental triggers, but do not provide smokers with just-in-time information about their risk of smoking lapse. Such risk information could be used to alert smokers to engage in relevant coping strategies including avoidance or use of quick acting pharmacotherapies (e.g. nicotine inhaler). The overarching premise of the proposed research is that prediction of lapse risk can be enhanced by using computer vision to analyze images of everyday life. Recent findings by our team suggest that environments associated with lapse risk can be detected in real-time by coupling deep learning-based object detection with an appropriate classification model. In preliminary research, images taken by smokers of smoking and nonsmoking environments (80 subjects, 2870 images) were used to train such a model, resulting in 77.5% accuracy (0.826 AUC) distinguishing these environments on a separate test set (16 subjects, 516 images). Encouraged by this preliminary finding, we propose to refine and scale up this system by a) creating a larger and more representative image database that includes a sample of everyday environments acquired from smokers (n=60) and b) testing novel, personalized approaches for increasing smoking environment classification accuracy and assessing smoking risk across three aims.
In Aim 1, we will improve smoking environment classification accuracy and prediction of smoking risk by re-training an existing deep learning model to recognize a broad set of smoking-related objects (e.g. packs of cigarettes/ashtrays).
In Aim 2, we will further improve performance by fitting personalized models that account for individual differences in preferred or typical smoking environments. In Exploratory Aim 3, we will predict craving and negative affect/stress using a similar approach. The proposed research represents an innovative and critical next step in the development of a system that identifies smoking risk and/or its antecedents in real-time to support a just-in-time adaptive intervention for smoking cessation.
The majority of smokers relapse within one month of quitting smoking. Environments (e.g. park) and their related objects (e.g. park bench) are associated with smoking and urges to smoke and can serve as triggers to lapse and relapse. The proposed research will use state-of-the-art computer vision and object detection to identify smoking risk environments and objects with the goal of eventually developing systems that alert smokers to such risks from everyday images acquired with cameras worn by smokers. Such a system can be incorporated into more comprehensive and effective smoking cessation programs.