Before modern medicine brought advanced blood and urine tests, doctors could only rely on taste buds to diagnose diabetes. The sweet taste in urine has always been the most distinctive pathological feature of diabetes, and the English word mellitus literally means "sweet". High levels of sugar in your body fluids means that something is wrong with your metabolism, either your cells aren't producing insulin anymore or insulin isn't working on them.

More than a decade ago, researchers identified a pathological feature with a more insidious relationship to diabetes. It is well known that the complications of diabetes include nerve damage, and nerve damage acting on the cardiovascular system can lead to arrhythmias. This change can be measured by both electrical and optical means. So it may not be long before doctors will be able to diagnose diabetes with a device worn on a patient's wrist, without the need to test blood and urine.

Turn the hands of time back to 2005, when only top athletes and seriously ill patients were using heart rhythm sensors. But today, one in five Americans wears a heart rhythm sensor. That's exactly why AI startup Cardiogram is trying to figure out the relationship between heart rate and diabetes. At the AAAI Artificial Intelligence Conference in New Orleans last Wednesday, Cardiogram presented the results of a study showing that the heart rhythm sensor and pedometer in the apple watch can be a good predictor of whether a person has diabetes. Provided, of course, that it is paired with the appropriate machine learning algorithms.

For some time now, Apple has been focusing on the changing role of its iconic wearable, the apple watch, which has gradually morphed from a "personal trainer" to a "personal doctor". Last November, Apple and health insurer Aetna entered into a partnership: as part of a pilot program to reduce healthcare costs, Apple donated more than 500,000 Apple Watch units. It also conducted a study with Stanford University to test the Apple Watch's accuracy in monitoring irregular heartbeats (which can potentially lead to strokes and heart attacks). The partnership with Cardiogram and UCSF is just another recent development for Apple in this regard. Cardiogram is a company founded by a former Google engineer and based in San Francisco.

Cardiogram has developed a free app to collect heart rate data from Apple Watch or similar sensors integrated with branded wearable devices such as Fitbit, Garmin and Android Wear. This app uses the same type of neural network as Google Speech to Text to analyze and interpret data such as heart rate and step count. As far as the data themselves are concerned, they are almost meaningless for disease detection. This is partly because of the bias in the data collected by the sensors; and partly because, to train a model that recognizes specific features, the data must first be labeled.

To figure out what a diabetic's heart rate looks like, you first have to find some diabetics. That's the value of UCSF, which launched a large heart disease research project called Health eHeart in 2013 to collect massive amounts of digital health data on 1 million people. As of mid-January, 196,000 people had participated in the study. Each participant completes a questionnaire about their known medical conditions, family history, medications, and blood test results. About 40,000 of them have linked their information to the Cardiogram app.

Brandon Ballinger, co-founder of Cardiogram, was formerly the technical lead for Google's speech recognition software. That's exactly how we get our data labeled," he said. From a medical perspective, each piece of labeled data corresponds to a dying patient, but it's a very small sample for what Internet companies do. "

As a result, Cardiogram has had to employ a number of techniques to train its neural network, DeepHeart, to help diagnose human diseases. One of these uses semi-supervised sequence learning technology, which was originally used to process text data such as reviews underneath products on the e-commerce platform Amazon. Only now Cardiogram uses it to process a range of heart rate measurements - about 4,000 measurements per week. After some mathematical transformation, this information will be turned into a single value that summarizes the heart rate variability. The researchers will correlate these values with the patient's labeled data, and then the formal training can begin.

The model trained by this method achieved an accuracy of 85% in identifying diabetic patients in the non-training group. This result is very close to one of Cardiogram's previous results. Last year, Cardiogram and UCSF released a joint study showing that DeepHeart can predict conditions like hypertension, sleep apnea, and atrial fibrillation with 80-90% accuracy based on a week's worth of data collected by the Apple Watch.

So, can Cardiogram's algorithm actually predict diabetes without directly measuring the sugar content of the blood? No one knows.

Mark Pletcher, one of the lead researchers on the Health eHeart project, said, "Clearly, diabetes is a cardiovascular disease, but there is no clear physiological link between it and heart rate variability. When you train a machine learning algorithm with data without understanding the underlying principles, you often get a result without understanding how it was arrived at. Frankly, it makes me nervous. We had a lot of internal discussions about whether this was a valid diagnostic tool or an irrelevant factor. At this time we have not reached a conclusion. "

It also raised the alarm of Eric Topol, a cardiologist and director of the Scripps Translational Science Institute, who once led the digital health arm of the NIH's multibillion-dollar Precision Medicine Initiative. He said, "Cardiogram's study is unconvincing and untenable in that it features both a biological black box and an algorithmic black box. This is, at best, a study based on hypothetical premises. "The premise of the hypothesis is that DeepHeart found a pathological signal for diabetes, but what it found could also be something else.

Ballinger was quick to refute such criticism and accusations. He believes that if the wearable tells the owner that he is at an increased risk of developing diabetes, that's when he can still get good treatment by going to a doctor for further diagnosis using traditional means. What if this black box could help us glimpse some doorways? But Ballinger also realizes that some prospective validation must be done to prove that AI does work. For example, screening out people who have not yet been diagnosed with diabetes and following them to see if they do end up with diabetes. He said Cardiogram is actively engaged in research in this area.

Cardiogram's apps for the Apple Watch and other wearable devices are currently free. But Cardiogram plans to add "user alerts" to the app later this year, sending notifications to patients the algorithm determines have atrial fibrillation, high blood pressure, sleep apnea and diabetes. Since the procedure has not been reviewed by the FDA at this time, it will not be used as a stand-alone diagnosis, but only as a recommendation. However, insurance companies might also pay for such advice if they think it will allow patients to be treated sooner and save on health insurance costs.

Cardiogram has a long way to go in the future, given the lack of sufficient convincingness in this study. Brennan Spiegel, a gastroenterologist and director of health services research, said, "The FDA will certainly review the accuracy, and in addition, there is a lack of data to support whether these wearable devices can actually change patient outcomes. Creating a technique is not the hardest part; the hard part is how to use those techniques to change the patient's behavior patterns. This has gone beyond computer science and come to the realm of behavior and sociology. "

But the Health eHeart project and Cardiogram's research show at least one thing: there is a strong demand for medical-grade measurement applications. Unfortunately, for a healthy person, these apps don't seem to do much more than send notifications. Thunderbolt.com Thunderbolt.com

via wired compiled by Thunderbird

1、Research What difficulties will enterprises face in their cloud computing practices
3、Redis MasterSlave Replication on Linux and SSDB MasterMaster Replication Environment Deployment Notes
4、Hacker Fundamentals of Tools ChapterAutomater
5、How to install and configure Postfix as a sendonly SMTP server on Ubuntu 1604

    已推荐到看一看 和朋友分享想法
    最多200字,当前共 发送