Ask people what they ate yesterday, and they will get it wrong. Not usually on purpose. Memory is bad, portions are hard to estimate, and snacks disappear from recollection entirely. Nearly everything public health knows about how people eat rests on that shaky foundation.
A team at Northwestern University has spent years building an alternative: a necklace that watches. Their system, HabitSense, combines a thermal sensor array, a small camera, and motion sensors, runs machine learning models on the device itself, and pings a paired smartwatch when it thinks the wearer is eating or smoking.
In a recent seven-day free-living test, the necklace generated 229 smoking alerts. Checked against video and the wearer's own answers, 209 of them, 91%, were correct.
Why Nobody Trusts a Food Diary
The problem the hardware is aimed at is measurement, not treatment. Self-reported dietary intake is the weakest link in nutrition science, and the same is true of self-reported smoking. People forget, round down, and reconstruct.
Passive sensing offers a different bargain: fewer questions, more data, and a record that does not depend on recall. It also raises an obvious objection: a camera hanging around your neck sees everything else, too.
The Northwestern group, led by Nabil Alshurafa's HABits Lab, built its answer into the device. HabitSense was described in a peer-reviewed paper in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, developed with input from clinician focus groups of 36 people and field feedback from 105 participants over 35 days.
An algorithm the team calls SECURE keeps the camera off until thermal sensing indicates that a health-risk behavior is beginning, so the device records activity rather than surroundings. On-device software then blurs the background while leaving the hands and face unobscured. Selective recording cut storage needs by 48% and extended battery life by 30%. In a seven-day study of 15 participants, the platform captured 768 hours of footage containing roughly 421 minutes of hand-to-mouth gestures.
An earlier necklace from the same lab, NeckSense, was built specifically for eating, using proximity, ambient light, and motion signals to pick up chewing, feeding gestures, and the lean-forward posture that tends to accompany a meal.
Seven Days, 780 Hours, and a Watch That Asked
The newest work tests whether the system can flag behavior as it happens, which is the prerequisite for any intervention that arrives in the moment.
Seventeen participants, nine in a smoking cohort and eight in an eating cohort, wore the necklace during waking hours for seven consecutive days. Whenever the algorithms predicted an event, the paired smartwatch asked a single question: "Are you smoking or eating?" The camera recorded video for independent visual confirmation.
The study collected 780.6 hours of sensor data covering 217 smoking episodes and 87 eating episodes. Of the 229 smoking predictions, 209 were true positives, and 20 were false alarms.
Two things should temper that number. It describes how often an alert was correct, not how many cigarettes the system caught, which is a different, equally important question. And the eating cohort was small, with eight participants and 87 episodes.
This work is posted as a preprint on the JMIR server and has not been peer-reviewed. The results may change before publication, if they are published at all.
Five Ways of Overeating
The reason to care about the plumbing is what the same tools have already produced. A study published in npj Digital Medicine in September 2025 used this approach on a question self-report cannot answer well: not how much people eat, but why a particular meal turns into overeating.
The SenseWhy study, which ran from 2018 to 2022, followed 65 adults with obesity in the Chicago area for two weeks each, collecting 2,302 meal-level observations, an average of 48 per person. Researchers manually labeled micromovements such as bites and chews from 6,343 hours of footage spanning 657 days and paired that with dietitian-administered 24-hour recalls and questionnaires delivered before and after meals.
Using those data, the team predicted overeating episodes with a mean area under the receiver operating characteristic curve of 0.86, then identified five distinct overeating phenotypes, each with a different trigger. Northwestern named them Take-out Feasting, Evening Restaurant Reveling, Evening Craving, Uncontrolled Pleasure Eating, and Stress-driven Evening Nibbling.
That taxonomy is the practical payoff. Four of the five cluster in the evening, but the trigger differs for each, and a stress-management strategy won't help someone whose problem is restaurant portions.
What a Necklace Still Cannot Do
None of this is a product, and none of it is treatment. These are research instruments used in studies, not devices anyone can buy, and no clinical trial has shown that wearing one changes what a person eats or whether they quit smoking.
The sample sizes are small by clinical standards. Detecting an eating episode is also not the same as knowing what was in it, and the systems described here do not measure calories or nutrients. The privacy engineering is thoughtful, but the underlying proposition is still that a camera records parts of a person's day, and that will not suit everyone.
Anyone worried about their eating patterns or trying to stop smoking has options available now that do have trial evidence behind them, including structured behavioral programs and, for smoking, medications and quitlines. A clinician is the place to start.
Key Questions Answered
What is HabitSense?
A research necklace built at Northwestern that uses thermal, camera and motion sensors with on-device machine learning to detect eating and smoking in real time. It is not a consumer product.
How accurate is it?
In a recent seven-day test, 91% of its smoking alerts were correct. That figure describes the precision of its alerts, not how many events it caught, and comes from a preprint that has not been peer-reviewed.
How does it handle privacy?
The camera activates only when thermal sensing indicates relevant behavior, and on-device software blurs the background while keeping hands and face visible.
What has this approach already shown?
A peer-reviewed study of 65 people with obesity identified five distinct overeating patterns, most concentrated in the evening but driven by different triggers.
Can it tell what someone is eating?
No. These systems detect that eating is happening. They do not measure calories, portions or nutrients.
Can I buy one?
No. These are research devices used in university studies, and no trial has yet shown that wearing one changes behavior.
