Researchers at app developer, Cardiogram, and the University of California, San Francisco, tested a deep neural network called DeepHeart that tells apart people with and without diabetes. The app was reported to predict early signs of diabetes with 85 percent accuracy.
For the study, Cardiogram relied on more than 200 million sensor measurements from 14,011 participants using an Android Wear or Apple Watch device with the Cardiogram installed. The data included heart rate, step count, and other activities that required physical activity.
This information was then used to train DeepHeart into identifying early signs of the disease. Researchers fed the neural network samples from people with and without diabetes, sleep apnea, hypertension, high cholesterols, and atrial fibrillation.
Prediabetes is difficult to pinpoint because traditional methods of detection require glucose-sensing hardware. However, an AI-based algorithm such as Cardiogram’s DeepHeart coupled with a wearable device such as Apple Watch has the potential to warn users when there’s an issue.
According to Cardiogram co-founder, Johnson Hsieh, regular deep learning algorithms require huge amounts of data, which are obtained from millions of categorized examples. In medicine, however, each label would represent a human life at risk, he added.
“To solve this challenge, researchers applied two semi-supervised deep learning techniques which made use of both labeled and unlabelled heart rate data to improve accuracy,” Hsieh stated.
The Cardiogram exec pointed to a link between diabetes and the body’s autonomic nervous systems, something that allows DeepHeart to identify the disease through heart measurements. In other words, those who suffer from prediabetes will have noticeable heart rate shifts.
The study was presented at the Thirty-Second AAAI Conference on Artificial Intelligence in New Orleans on Wednesday.
Cardiogram plans to implement DeepHeart directly into the Cardiogram app, which will alert Apple Watch users if early signs of diabetes are detected.
Image Source: Flickr