ML for Weight Loss on WW (WeightWatchers) App

Wellness is big business. But while the total market value of the weight loss industry has recently grown to all-time highs, the growing role of machine learning technology in weight loss products and services could portend a new surge for this sector. That’s because machine learning offers the opportunity for far greater personalization at scale, which may prove especially valuable in the weight loss category.

Specifically, one app we’re looking at today is WW. WW, formerly WeightWatchers, is one of the oldest and most successful brands in the weight loss industry. It was founded nearly 60 years ago, but has evolved to keep up with shifts in technology in impressive fashion. At its core, WW is about building a community support system around making healthy food choices, guided by a food point system. Users are meant to track their daily dietary intake, using WW to calculate the total point value of their consumption, aiming to stay within certain point parameters. Each week, users weigh in, and consult with experts and peers about their progress.

As you can imagine, the way this offering has manifested over the past 6 decades has changed significantly, as technology and culture have changed what consumers expect and what is possible. For example, while WW still reports that nearly 1.5 million users attend weekly check-ins/counseling sessions, the app also now includes a robust digital community forum and supports at-home weekly weigh-ins. Another significant change has to do with the manner in which users track their consumption and manage their points. Whereas this was once done manually, using pocket-dictionary-like books full of foods and point value equivalents, users can now simply query the enormous catalogue in-app, or- simpler still- use their smartphone to scan the barcode of any commercial food item.

The opportunities for machine learning in this app seem nearly limitless. One of the most difficult elements of any good machine learning application is having high quality, consistent data that can be used both to train a model and then to rely on that model to make future predictions. When it comes to consumer behavior, passive tracking is becoming more popular and viable (such as with wearable technology, and other sensors and trackers), but consumption is one of many areas where manual logging is still required of users and a tall order in most contexts. In the case of WW, however, millions of users are already proactively tracking every food and beverage they consume each day, providing a valuable set of input data that can be used to predict several important outputs. 

First and foremost, the most coveted output of a WW user is their weight. While the overall WW program and points structure are designed to provide consistent, week-over-week weight loss, machine learning could provide greater personalization to this by using past consumption and weight loss to predict how future consumption will impact future weight loss for a particular user. After some time on the app, a user could start to expect not just a daily or weekly point target, but recommendations about how specific foods seem to be impacting the users’ weight loss journey, so they could respond accordingly. 

One of the things that makes this such a strong candidate for a machine learning model is that each data point (a food or drink consumed) is actually made up of a plethora of other, very standardized and relevant data. For instance, every time a user logs that they ate an egg, the model can learn not just that the user ate an egg, but that they also consumed roughly: 78 calories, 5 grams of fat, 1.6 grams of saturated fat, 6 grams of protein, etc. Each of these data points further paints a picture of the dietary inputs that are contributing to the user’s weight.

But weight is not the only output target that the WW app tracks. During digital weekly check-ins, users are asked to reflect on other elements of their wellbeing. While some of this involves providing open-ended narration, much of it also is a quick-capture of the user’s feedback from a structured set of options (i.e., selecting an answer on a 1-5 scale, or from a finite multiple choice list). This kind of standard, tabular, and consistent data is ripe for machine learning modeling. Again, the app could begin to predict various outputs based upon the day-to-day consumption inputs of the user.

Key to a quality machine learning model is having input data that is reasonably and reliably predictive of the target output. The science of weight loss is extremely clear: how much one consumes is incredibly predictive of changes in one’s weight. So on that front, we should be quite optimistic that consistent tracking of consumption would make for strong input data for a model meant to predict future weight gain or loss. That said, it is likely true that there are other variables that contribute to weight fluctuations, and we can use machine learning to help suss out some of those relationships, too. That’s because the WW app already integrates with wearable technologies, such as the Apple Watch and Fitbit, and other tools to help capture even greater additional input data that may impact weight. Things such as workouts recorded on an Apple Watch, stand-time recorded on a Fitbit, or even the amount of mindfulness sessions one does in a day, or one’s average amount of screen time, might all have some bearing on an individual’s weight loss goals. Machine learning could help identify those patterns and nudge users accordingly.

We’ve undoubtedly just scratched the surface, but it is clear that there exists a world of new opportunities for an app like WW to harness the power of machine learning to provide users even more tools on their journey to lose weight. Given the immense amount of data, a company like WW might even be especially well suited to pioneering entirely new insights about the relationships between diet and various health outcomes, and it will be exciting to see how the company continues to evolve to take advantage of new tech.