The combined models identified smoking environments with 76.5% accuracy.
A smoker’s favorite bench, or a spot in the park — do they spur cravings? And can artificial intelligence (AI) predict when the urges will hit?
A deep-learning approach can trigger just-in-time adaptive smoking cessation, optimize a smoker’s environment during quit attempts and study environmental correlates of other behaviors, according to new research published in JAMA Network Open. The AI model identified environments associated with smoking with 76.5% accuracy.
A smoker’s cravings increase when they are in certain environments, which could provoke lapses during a quit attempt.
“Understanding how the external environment affects behaviors and symptoms of interest may inform environment-based interventions and therapeutic environmental modifications,” the study authors wrote.
Rather than a typical clinical approach of asking smokers to identify environments they smoke in and telling them to avoid the area or cope in a different way, the research team used AI to identify specific environmental triggers.
Researchers recruited 106 participants from Durham, North Carolina and 63 from Pittsburgh, Pennsylvania from 2010 to 2016. Participants were between 18 and 55 years old and smoked at least five cigarettes per day for at least one year. The participants were ambulatory, not ill during the study and were not planning to quit smoking during the study period.
Each participant took up to four pictures of daily smoking environment and up to four daily nonsmoking environments.
A smoking environment satisfied two of the following:
Nonsmoking environments satisfied two of the following:
Participants took two pictures as they approached the environment and two as they were within it.
A subset of 37 participants viewed eight images of standard environments and reported the craving associated with each one on an eight-point scale.
Researchers trained the AI model to identify the images as either smoking or nonsmoking environments. The first part of the model gave information about image content based on 1,000 categories — common objects or locations, such as a patio, trash can, library, desk and printer. The second part of the model relates information about the presence or absence of the features to the probability that the image depicts a smoking or nonsmoking environment, according to the study.
Three validation schemes evaluated the performance within and between Durham and Pittsburgh. In the first, the model was developed and validated using images from Durham and applied to the Pittsburgh images for second validation. The second scheme was reversed. In the third, the model was developed and validated using the images from both geographic locations.
The mean area under the operator curve in the was 0.84 with a 76.5% accuracy. Test performance was higher on the Durham images regardless of which training set was used. Including the Durham images improved results compared with training with just the Pittsburgh images.
Three of four experts’ performance was above the sensitivity and specificity curve for the models trained with the first and third model. But the differences were only statistically significant for one expert, who outperformed the Pittsburgh-trained model and the final model.
The research team sees a just-in-time adaptive intervention where images from a wearable camera or smart glasses are assessed on an ongoing basis to quantify smoking risk and trigger an intervention when the risk is high.
The findings can inform environment-based interventions. For example, images of a destination could be analyzed before a smoker visits to see whether the environment could increase craving.
The AI model could also support therapeutic environmental modifications to promote healthy behaviors. Images collected during a failed quit attempt could be analyzed to pinpoint environmental factors associated with relapse.
“Working together with a clinician, the smoker might then remove these factors from their personal environments before their next quit attempt to increase the chance of success,” the authors wrote.
Get the best insights in digital health directly to your inbox.