A recovering cigarette smoker finds herself in a place that triggers strong associations with smoking and is sorely tempted to light up. But what if artificial intelligence could become an effective partner with would-be quitters, identifying environments predictive of smoking and intervening in the nick of time to nip that craving in the bud?
A study has found a deep learning approach may be able to do that, recognizing locations predictive of smoking and triggering "just-in-time adaptive cessation interventions." It could also help optimize smokers' environments during cessation attempts and, more broadly, analyze the environmental stimuli of other behaviors that need modification.
The deep-learning approach successfully differentiated environments that participants designated as smoking or nonsmoking with a mean area under the curve (AUC) of 0.840 (standard deviation, 0.024) (accuracy 76.5%; standard deviation 1.6%), a performance comparable to that of human smoking-cessation experts.
"The findings suggest that objects and settings found in images of daily life can be used to identify environments associated with smoking, which may in turn be an effective proxy for craving and smoking risk," write lead author Matthew M. Engelhard, MD, PhD, a senior research associate in Duke University's Department of Psychiatry and Behavioral Sciences in Durham, North Carolina, and colleagues. The study was published online August 2 in JAMA Network Open.
Innovative approaches are very much needed; according to the Centers for Disease Control and Prevention, the best cessation interventions achieve 6-month abstinence rates of less than 20%.
Study Details
The cross-sectional study recruited 169 smokers from Durham, North Carolina and Pittsburgh, Pennsylvania (ages 18-55) and had them photograph 4902 images of smoking (n = 2457; 50.1%) and nonsmoking (n = 2445; 49.9%) locations from 2010 to 2016. All participants were active smokers lighting up five or more cigarettes a day for at least 1 year.
Smoking locations were frequently visited sites where participants often smoked or found it hard not to, whereas nonsmoking sites were those often frequented but not smoked in. Direct smoking prompts and cues such as cigarettes or lighters were not photographed.
These images were then used to develop "a probabilistic classifier" for predicting participants' smoking or nonsmoking locations, correlating settings in daily environments with smoking patterns. The classifier "combines a deep convolutional neural network with an interpretable logistic regression model and was trained and evaluated via nested cross-validation with participant-wise partitions (ie, out-of-sample prediction)," the researchers explain.
At the real-life level, Engelhard told Medscape Medical News, "there are very nice machine learning frameworks designed to run models like this on mobile hardware — an example being TensorFlow Lite — so it can definitely be implemented." His group is currently working to optimize the model for mobile devices.
"So, for example, you'll be able to take a picture with your mobile phone and run the picture through the model right there on the phone to estimate smoking risk associated with the picture," Engelhard said. Your smartphone could then remind you of specific strategies for parrying urges, connect you with a designated support person, or perhaps recommend taking a nicotine lozenge or similar product, he explained.
A total of 3386 images were available from Durham and 1516 from Pittsburgh. Images were evenly split between the two classes, with 2457 smoking images (50.1%) and 2445 nonsmoking images (49.9%).
When applied to unfamiliar, previously unseen environments, the model's predictions strongly correlated with participant-reported craving associated with those environments. "This provides preliminary evidence that environmental patterns associated with smoking may confer risk wherever they are encountered," the authors write.
The top environmental locations or features for smoking were a patio, a moving van, a park bench, a space heater, and a pool table. This category also included, surprisingly, a sundial, a mosquito net, and a snake-rail fence. Less surprisingly, top nonsmoking environments ranged from a library, grocery store, and photocopier to a shoe shop and a church.
A model trained only with Durham images effectively classified images from Pittsburgh, for an AUC of 0.757 (accuracy 69.2%), and a model trained only with Pittsburgh images effectively classified images from Durham, for an AUC of 0.821 (accuracy 75%). Those findings point to good generalizability between geographic areas.
Only one outside expert outperformed the classifier at a statistically significant level (P = .05). In addition, the median self-reported craving of participant significantly correlated with model-predicted smoking environment status (ρ = .894, P = .003)
In another part of the image study, 25 randomly selected participants took 732 photographs, which were classified by four smoking cessation experts not involved in the research. These classified each image (yes or no) on the basis of the following question: "Would you warn a smoker that this is an environment in which they might smoke or be tempted to smoke?"
According to Engelhard, the experts "were thrilled to see the algorithm perform so well, and they see this as an important breakthrough in helping smokers successfully quit. This model would not replace clinicians — it is just a tool to help patients and providers work together more effectively to help the patients quit successfully."
The authors call for more systematic study of how daily environments promote smoking behaviors — "information that can be leveraged to support more effective quitting."
Psychiatrist John B. Torous, MD, director of digital psychiatry at Beth Israel Deaconess Medical Center in Boston, Massachusetts, told Medscape Medical News the results from the large two-site analysis are "impressive and positive," as science moves toward just-in-time adaptive interventions (JITAIs) "This is the type of research that would provide clues as to what is the right time to deploy these JITAIs and just when they would be most effective," said Torous, who was not involved in the study.
He also raised the issue of privacy and ethical concerns, a constant concern with emerging digital applications. "If we're taking photographs of various locations, and there are bystanders picked up, for example, what systems and mechanisms are in place to keep this information private, safe, and protected?"
Last year, Medscape Medical News reported results of a global survey showing that smoking behavior is deeply embedded in the daily rituals and routines for waking, eating, drinking, and socializing, and that underscores the need to focus on the nonphysiologic aspects of tobacco addiction.
Looking ahead, the authors say the results suggest a potential framework for predicting the impact of daily environments on other target behaviors or symptoms, such as mood disorders, attention deficit hyperactivity disorder, obesogenic behaviors, and allergen-induced asthma attacks. "Understanding how the external environment affects behaviors or symptoms of interest may inform environment-based interventions and therapeutic environmental modifications," the authors write.
They plan to explore the possible benefit of personalization of the model by identifying settings associated with smoking on an individual basis.
This work was funded by grants from the National Institute on Drug Abuse. Study authors Engelhard, Jason Oliver, Lawrence Carin, and Joseph McClernon report a pending patent for Predicting Real-time Behavioral Risk Using Everyday Images related to the work in this study. Torous has disclosed no relevant financial relationships.
JAMA Netw Open. Published online August 2, 2019. Full text
For more news, follow Medscape on Facebook, Twitter, Instagram, and YouTube
Medscape Medical News © 2019
Cite this: Deep Learning AI May Help SmokersKick the Habit - Medscape - Aug 06, 2019.
Comments