Predicting Biological Age for a HealthTech App

Humanity is a health tech startup dedicated to extending the healthspan of every person on earth, interested in applying AI for aging. The Omdena team built predictive aging models for Humanity´s app.

“Since launching the app worldwide at the end of 2021, Humanity Inc gained 75 000 sign-ups with 50 000 actively engaged users! 40 days after signing up monitored health actions of users increased 15%. This gives 0.49 years per active user so far, validated against published and peer reviewed scientific aging models.

And that’s only the beginning! We are glad that at Omdena, we can be part of this healthy life extending revolution.”

The problem

Humanity will steadily grow the userbase to millions of users over a coming couple of years. We need to build a system that takes in the user actions that we are monitoring on one side (activity rates, sleep, meditation, diet, etc.) and uses the ongoing increases or decreases in the user’s Rate of Aging measure (a separate model based on third party longitudinal data that includes actual health outcomes) to decide which actions were most effective and in what combinations and when.

The system then also needs to be able to match across users with similar attributes to use the insights and weightings set for one user to affect the weightings given to actions and combination of activities to another user.  You can think of this as very similar to the way the Waze traffic app works. It uses the real-world traffic routes and timings of people driving similar routes or parts of routes just a few minutes ahead of you to allow you to know the fastest route to take from your current location and intended destination and approximately how long it will take you. This system will effectively do that but for helping the person navigate towards a healthier longer life.

Current metastudies on the effects seen from introducing certain new actions/interventions to a person’s daily actions prove that they can have a large effect on the person’s rate of aging, but they do not allow the personalization and combinatorial nature of real-life to be modeled. Put simply, adding two positive actions that work for most people (proven by separate research studies) does not necessarily bring a positive result for the particular person and, even worse, may cause a negative effect. Thus Humanity has built a system to capture that data, and Omdena trained the models.

Humanity clarified the methods and exact list of all biomarkers and actions being monitored currently and which ones can be added in the near future.

The project outcomes

In this project, Humanity and the Omdena team compressed high throughput markers such as activity and other lifestyle action data from the user using unsupervised techniques and then used supervised techniques to develop weighted algorithms predictive of the biological age outcome. There is a focus on lifestyle actions and their effect on the outcomes and stratifying and projecting those effects across the userbase matched by an individual’s wider attributes (e.g., diet, weight, socio-economic status). The outcomes for the project are:

• Code for compression of high throughput lifestyle action data
• Code for model training and evaluation of models on biological age outcomes
• Code and measures of the trained model in independent testing dataset
• Final trained models, assessing the value of each action independently and in combination, stratified by individual situation.
• Release of final trained models to the userbase with projection onto actions observable directly by humanity