Research detects smoking via sensor data

Anyone who’s tried to quit smoking knows that the path to success is fraught with challenges. Every day you get up with the best of intentions, but as the day progresses, there are tests.

Finding a way to keep that first lapse from happening would be a huge step towards helping people to remain tobacco-abstinent. And the importance of giving up tobacco cannot be underestimated. Smoking is the leading cause of preventable death in the United States, according to the Centers for Disease Control. One in five deaths, or about 1,300 deaths a day, can be attributed to smoking or exposure to second-hand smoke. For every person who dies from smoking, at least 30 have a serious, smoke-related illness.

If you’re someone who’s trying to quit smoking, the things once associated with smoking – the morning coffee, eating, riding in the car, having a beer with friends – become potential triggers for a lapse. Research has shown that, once that first puff is taken, a person attempting to quit is far more likely to begin smoking again.

What if a newly-abstinent smoker could wear a sensor that would detect when they have that first puff? What if it was possible to intervene before that puff turned into a full relapse into smoking? The Center of Excellence for Mobile Sensor Data-to-Knowledge (MD2K) has made remarkable progress toward this goal in its first year of research.

A computational model, called puffMarker, was found to be quite accurate in detective a first lapse when tested on 61 newly-abstinent smokers who wore MD2K sensors for 3 days following a quit attempt. The approach was developed by a team led by Nazir Saleheen, a Ph.D. student at the University of Memphis, working with MD2K director Dr. Santosh Kumar.

The puffMarker model fulfills a long-standing need of the smoking research community. The work was recently published in a paper presented at the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp).

MD2K is an NIH-funded center tasked with developing the software, tools, and science to use mobile sensor data to improve health. A second area of research is developing ways to use sensor data to reduce hospital readmissions among congestive heart failure patients.

The puffMarker model uses data collected from wearable sensors:  a RIP (respiratory inductive plethysmography) sensor that captures breathing patterns and wrist sensors that capture movement from accelerometers and gyroscopes.

Wrist sensors can detect when a hand moves to the mouth or face, but that can mean anything: eating, sneezing, or scratching. The challenge MD2K sought to address was how to distinguish the gesture of smoking from all other similar gestures that naturally occur during daily life.

By adding a RIP sensor, researchers were able to identify a combination of respiration and gesture: a hand moves to the mouth, followed by a deep inhalation.

“We observe that during smoking, the hand comes to the mouth and is immediately followed by a deep inhalation,” the researchers said in the puffMarker paper. “During walking, the hand is downwards with a pendulum-like movement, and respiration is faster. During eating (cereal), the hand comes at the mouth; however deep inhalation, observed during smoking, is absent in such activities.”

The MD2K team tested the model in the lab using data from regular smokers, and found it to be 96.9% accurate with a false positive rate of 1.1%. Then they applied to data collected in a smoking cessation study with 61 participants who were attempting to quit. Each participant wore the sensors 1 day while smoking and then 3 days after quitting. Of those, 33 lapsed within the 3 days after quitting, which was verified by a carbon monoxide monitor. The puffMarker model was applied to 32 of the 33 who lapsed (1 set of results had to be discarded due to data loss). The model was able to detect the first lapse in 28 of the 32 participants. When tested on those who abstained during all 3 days, the false positive rate on those abstinent smokers was 1 episode every 6 days.

The paper cited several ways in which the model could be improved for future research, including personalization that calibrates the model using data from each participant to improve accuracy; expanding the model to detect other behaviors; finding a way to obtain accurate results from just one wrist sensor; adjusting the model for cigars, e-cigarettes and hookah, and adjusting the model to detect isolated puffs or when several people share a cigarette.

“Our work opens up a very rich area of research for discovering just-in-time interventions that can be triggered from predictors detected by sensors such as GPS, smart eyeglasses, electronic and social media, and physiological sensors,” the paper concluded.

Copyright © 2017 MD2K. MD2K is supported by the National Institutes of Health Big Data to Knowledge Initiative (Grant #1U54EB020404)

Team: Cornell Tech, GA Tech, Harvard, U. Memphis, Northwestern, Ohio State, Open mHealth, UCLA, UCSD, UCSF, UMass, U. Michigan, Utah, WVU