Applying Artificial Intelligence to Help People Quit Smoking: Early Results

Published in

Towards Data Science

9 min readMay 30, 2017

Notes: All data reported is from actual users, sanitized to protect their personal information. Our data collection and reporting systems are designed to keep our users’ personal information private.

tl;dr we are going to rid the world of smoking, its challenging but we have a plan, and need your help. Sign up at cue.kiwi.ai

When was the last time you thought about quitting smoking? If you are like any of the ~1 billion smokers in the world it was probably this year. If you happen to have a smartwatch, we might have a way to justify its purchase, our app Cue is your champion to reduce smoking

A key part to maintaining any habit is consistency, and ease of use. If changing my eating habits requires me to carry a clipboard around and log everything I eat, then I’m changing two habits 1) eating healthier 2) logging the eating … which is considerably difficult in a world where when Google takes longer to search for something its 0.59 seconds

Illustrative Typical Behaviour Change Apps/Products

When we set out to make software to reduce smoking, we had two key rules in mind 1) it cannot make the user do more work 2) it must deliver a benefit beyond the novelty phase … we were not going make up vanity metrics

1) Machine Learning to deliver an extremely personalized experience, that drives actual benefit

We have applied similar thinking to tracking cigarettes; we have developed an algorithm to automatically detect when you smoke. Our app is your champion to nudge you on the path to quit by making small yet consistent improvements to the average time between cigarettes

Background on Kiwi.ai and Cue:

Four years ago we made our first prototype for tracking when you smoke, here is a link to the open source project; the technology and market adoption wasn’t there then, and we weren’t going to build a custom wearable for smoking … so we waited

In that time we built and licensed motion recognition software that used embedded neural networks to track movement with high accuracy levels with low power consumption. Since then we have supported over 20 customers including large brands across large technology companies such as IBM, OEM device manufacturers, robotics companies and healthcare providers

We were often thinking about use cases that would materially solve a problem for the consumer, to make their health really better, to save them money, to make them happier. Perhaps its because we are still in early stages of wearables, but still we needed to give the consumer a reason to wear their wearables.

Smoking Reduction

Smoking is a real problem. It kills people, costs the healthcare system a ton and is just all round terrible for all the players except those who sell cigarettes

Its also a ridiculously difficult problem to solve, otherwise we still wouldn’t have a billion smokers around the world, that try to quit at least once a year with little success

Kiwi is on a mission to make the lives of at least a billion people happier and healthier

Its a product that delivers tangible health and financial benefits e.g. if it works, you know, you feel+look better and everyone (including your loved ones, bank account, doctor, … is happier) and is something that can be delivered 10x better using a watch app as opposed to a phone

It also helped that two people on our team were smokers, so we had the first guinea pigs to test out the earliest versions of our software. After speaking to hundreds of customers (which is easy because they are usually hanging out outside buildings), we continued to refine the design, features and user experience

We recently ran an initial user group, based on our target demographic but people we did not know. We find that our friends worry about our feelings too much

The first hypothesis we tested was if people would actually wait a bit longer than usual to smoke their next cigarette. With version 1 of our app (automatic tracking at 99%, thanks to our work in motion recognition), we sampled a population group over 30 days

Here is a quick overview of the data we have put together thus far:

key points: people will wait longer than usual, they often smoke at 3pm, with wide variations in daily smoking patterns (after all life is unpredictable), the general trend is consistently towards fewer cigarettes if an extremely personalized plan to reduce smoking is deployed

Hypothesis 1: Do People Wait for Suggested Cigarette Time?Before Suggested Time: 0359
After Suggested Time: 1049
Total Cigarettes : 1408Answer so far: YES

Most Common Time To Smoke (3PM “coffee break”)

Our Market Test User Base : Wide Variations on a Daily Basis

Up to 50% Reduction Across 3 month Period Over a Testing Sample User Base

Validating Our Hypothesis

This is where we need your help; we have tested locally with a small group in Toronto who each have had their own success of up to 50% reduction.

We now want to roll out our app to as large a group as we can manage; people to test further, and help us further refine our product to kick smoking in the butt

The habit altering part of the app encourages you to wait a few minutes longer for your next smoke by receiving points, these points get you rewards at Starbucks/Uber, win-win! We’ve requested some of the best experts on smoking cessation to find gaps in our plan, and they have all encouraged us to pursue this project. But there is nothing like 10,000 healthier users to confirm a hypothesis, so we need your help

If you smoke, please sign up or otherwise please help by sharing with your friends who smoke. You can help make a big difference in someone’s life with a couple of clicks

Technical Aspects

The remainder of this post talks about the machine learning aspects of our product, involving

Identifying each person’s unique movement patterns & discerning smoking from all other actions [Auto Classification]
Predicting when a person is going to smoke based on individual and aggregate data

Auto-classification: Best User Experience

We have found the best user experience is one where the user does not have to do anything and still feels like they are winning. To do this with smoking we needed to come up with a method to automatically detect when a user is smoking. We do this by using our kiwi toolkit in with motion recognition, much like how steps are counted we detect hand to mouth movements to define when a person is smoking

An important note to this is all our classification algorithms run on the watch; so once the app is download no connection to your phone or internet is needed to make it work!

A neural network is used to separate movements that may overlap such as lifting objects, walking with your hand raised or talking enthusiastically

Here is a fancy plot which shows some of this separation, in a Linear Discriminant Analysis:

How Smoking Separates From Walking and Other Movements

Predicting When a Person is Likely to Smoke

In order to best help people nudge the time between cigarettes to longer and longer periods; a prediction needs to made for when your anticipated next cigarette will be, from here we nudge slightly to increase the time by a little bit to keep the effects of the behaviour change subtle

Here is a plot of daily occurrence of when people smoke over a week:

See how there are breaks each day and Saturday and Sunday are a bit later than weekdays

Prediction and Why This Is Valuable

Similar to being able to automatically detect when a person is smoking, a method can be applied to predict when a user is smoking in the day, like most model problems the challenge is how to get clean data and ensure you have a benchmark to correlate with. For instance, a Convolutional Neural Network can be used for photos, where different RGB signals can be separated to define the difference between labelled cat and dog photos. How can this be applied to just a few data points in a day such as smoking times?

Feature Separation For Predicting a Person’s Next Smoke

What we do is generate features to generalize across our entire user base to find distinguishing pieces of information, the plot above shows a sample of three features:

Feature 1: time between current and last cig
Feature 2: current day of week, format: [0..6] -> Mon..Sun
Feature 3: current time of day, format: HHMM

Using these features we are able to come up with a prediction to the nearest 30 mins or 60 mins; with this we can help users pre-empt the urge of when to smoke next by going for a walk, playing a game or grabbing a coffee on us, more to come!