Fitbit and Apple Watch can help predict diabetes, says report
DeepHeart: Fitbit and Apple Watch can help predict diabetes risk

Fitbit and Apple Watch can help predict diabetes, says report

Smart watches just got smarter, according to a new study of the use of wearables to predict the risk of medical conditions, including diabetes, high cholesterol, and high blood pressure.

An AI neural network, known as DeepHeart, is the brains behind the breakthrough.

Research from digital heart-rate tracking company Cardiogram has revealed the latent potential in consumer heart rate trackers, such as those found in Fitbit and Apple Watch devices, to detect signs of cardiovascular illnesses. They presented their findings at this week’s AAAI Conference on Artificial Intelligence in New Orleans.

By analysing the relationship between the heart rate and step counting data recorded by compatible wearables, Cardiogram was able to predict whether the participants had diabetes, with 85 percent accuracy.

Alongside diabetes risk, the research, carried out in partnership with the University of California, sought to train the company’s DeepHeart neural network to predict high cholesterol, high blood pressure and sleep apnea.

The study compared two semi-supervised training methods, sequence learning and heuristic pretraining, and successfully demonstrated that these methods can outperform traditional hand-engineered biomarkers.

The DeepHeart neural net

Existing (and widely used) predictive models rely on very small amounts of positive labels (which represents a ‘human life at risk’). However, readily available wearables such as Apple Watch, Fitbit, and Android Wear devices, benefit from trillions of unlabelled data points – including rich signals such as resting heart rate and heart rate variation, which correlate with many health conditions. As an individual develops diabetes, their heart rate pattern changes, due to the heart’s link with the pancreas, via the autonomic nervous system.

Utilising consumer heart rate trackers offers a rich vein of data with which to train a neural network. This kind of AI thrives on huge quantities of information, as seen in natural language processing algorithms from the likes of Amazon and Google.

The research was not straightforward, however. Tracking company Cardiogram had to overcome several challenges presented by consumer-grade devices, including sensor error, variations in the rate of measurement, and daily activities confusing the data.

The company is now planning to launch new features within its app for iOS and Android, incorporating DeepHeart.

Internet of Business says…

We’ve touched on the wealth of data that healthcare providers could potentially tap into when it comes to wearables, such as the KardiaBand. This example requires supplementary hardware, however. With DeepHeart’s intelligent use of neural network methods, they have opened the door to healthcare professionals being able to make use of the persistent monitoring capabilities of consumer wearables.

With an estimated 100 million-plus US adults now living with prediabetes or diabetes, many of whom aren’t aware of having the condition, Cardiogram’s study has significant practical implications. This is magnified by the fact that one-in-five Americans own a heart rate sensor today, so the infrastructure is already there to deploy DeepHeart’s technology quickly. With rumours that Apple is considering including a glucose monitor in it’s next smart watch, the scope for using data from consumer wearables is set to grow even further still.

The likely determining factor in adoption will be the rate of deployment. Hospitals are typically slow to adopt new AI technologies because the cost of errors is so high.

A word of warning, too, we’ve already seen the danger of using ‘black box’ AI systems in our finance and justice systems – the dangers of using similarly opaque methods in healthcare are just as acute.

Read more: Police need AI help with surge in evidential data