Meet Flo – The First Period & Ovulation Tracker that Uses Neural Networks

Artificial Intelligence Case Study

Innovation is quickly entering the healthcare domain, disrupting it for good. To date, startups focusing on women’s health alone have raised over  $1.1B [1], and are growing in quantity and quality of services offered.

Some years ago, period and fertility tracking consisted in taking record of one’s menstruation days online or offline. Nowadays, with the help of new  technology, women can track a variety of symptoms, get advice on conceiving, explore pregnancy insights, and even manage irregular periods.

A major breakthrough in this domain has recently been done by Flo Period Tracker – the first period tracking app to publicly announce using artificial  intelligence for improving cycle predictions. Flo became the most downloaded app worldwide in its category within months after introducing neural  networks to its prediction algorithm. It’s no surprise considering around 30% of women around the globe face the challenge of irregular periods. Let’s dig into how the technology works.

Flo’s neural network became possible thanks to all the data its users have supplied for the app. When Flo decided to partner with a data science  company InData Labs to implement the neural network there were more than 450 GB of useful information stored in their database. Every day,  women manually log around 1.4 M new data points including cycle history, ovulation and pregnancy tests results, age, height, weight, lifestyle  statistics about sleep, activity, and nutrition. In addition, more data comes from wearable devices like Fitbit & Apple Watch.

A formula created for an “average woman” cannot compete with neural networks when we deal with the volume, velocity, and variety of data.

The beauty of Flo’s neural network is that it knows overall statistics and common rules that work for the 4+ million women using the app, but it still  treats every woman as a unique person with unique body characteristics and lifestyle.

Individual accurate predictions about menstrual cycle length and fertility windows are likely to help conceive faster when getting pregnant is the goal.  Personalized health insights provided by Flo keep women aware and in charge of their health at all stages of life – the start of menstruation, before,  during and after pregnancy, as well as menopause. This helps to reduce unnecessary medical intervention and cut healthcare expenses.

The more details a user supplies for the neural network to analyze, the better predictions she gets. This allows women to discover their body and learn  about meaningful connections between a variety of symptoms and activities, rewarding them for all the sensitive data that they entrust to the app.

How the artificial brain works

The basic idea behind a neural network is to simulate a large number of densely interconnected brain cells inside a computer so that you can get it to  learn, recognize patterns, and make decisions in a human-like way.

Neural networks consist of millions of artificial neurons called units, which can be divided into 3 types: input units receive various forms of  information from the outside world, output units represent neural network’s response to the information it’s learned, and hidden units connect input and output layers and form the most sophisticated part of the ‘artificial brain’.

Flo’s neural network has 442 input units that receive a variety of features engineered using the information women have supplied to the app. The  combination of features is unique for every woman and each feature has a different impact on predictions. The hard job of feature prioritization and  cycle length prediction is done by the neural network. It continuously learns by comparing its predictions to true results and fine-tunes interconnections between its input and output units over time. The output layer is represented by a single unit that transmits the predicted number of  days in the cycle. The current version of Flo’s neural network can improve irregular cycle predictions by up to 54.2% depending on quality and  quantity of data, with prediction error reduced from 5.6 to 2.6 days. Adding new features promises to produce even better results.

Flo team found a source of new features in the uniqueness of its users. Women manually log their mood, physical inner activity, symptoms like  headache, fatigue or acne, which sometimes form a stable pattern that repeats on certain days of a cycle. These unique repeatable patterns are so  individual that no human can create enough rules to capture them all, but they may be so evident and stable for a particular woman that by analyzing them the neural network can make a better prediction. For that reason, the data science team working on the project developed a machine learning  algorithm that can capture the unique menstrual cycle patterns for every woman.

Technically, this is realized through a two-step process. At the first step, unique patterns are recognized by individual-level machine learning models.  At the second step, the patterns are transformed into features for the neural network. Thus, an output from one algorithm becomes an additional feature for the neural network.

While working on its neural network , Flo with the assistance of InData Labs is undertaking research projects in the field of female health. The  company has just finished research on how resting heart rate varies across the menstrual cycle. The remarkable thing about the research is that it was  fully based on data supplied to Flo by its users with heart rate statistics coming from wearable sensor technologies like Apple Watch and Fitbit Charge  HR .

Those who wear health trackers like Apple Watch or FitBit are not limited to early adopters anymore. Thousands of Flo users have already integrated  the wearable devices with the app to be more aware of their health, and the number is growing. All this enables the company to conduct health  research at large scale and continuously improve its neural network predictions feeding it with new types of health data.

[1] https://www.cbinsights.com/blog/femtech-market-map/