[ad_1]
Deep medicine: how artificial intelligence can make human health again
by Eric Topol
The Apple Watch has caused a radical change in the public acceptance of biometric monitoring. Granted, we've had step counters, heart rate monitors and sleep monitors for years, but the Apple Watch has made it trendy and cool. In Deep medicine, author Eric Topol, examines how recent advances in AI and machine learning techniques can be leveraged to bring the (at least American) health system out of its dark era and create a more efficient system and more efficient, better serving both his doctors and his patients. In the excerpt below, Topol examines the efforts made by the startup AliveCor and the Mayo Clinic to integrate the functionality of an ECG into a wristwatch-sized device, without – and that's the important part – generating false positive potentially life-threatening results.
In February 2016, a small start-up company called AliveCor hired Frank Petterson and Simon Prakash, two Googlers with AI expertise, to turn their electrocardiogram (ECG) activity into a smartphone. The company was in trouble. They had developed the first single-pass ECG smartphone app and in 2015 they were even able to display the ECG on an Apple Watch. The application had a "wow factor" but otherwise seemed to have little practical value. The company was facing an existential threat, despite the significant venture capital investment of Khosla Ventures and others.
But Petterson, Prakash and their team of three other talents only in the field of AI had an ambitious double mission. One of the goals was to develop an algorithm to passively detect a heart rhythm disorder, the other to determine the level of potassium in the blood, simply from the ECG captured by the watch. It was not a crazy idea, since AliveCor had just hired. Petterson, vice president of engineering at AliveCor, is tall, blue-eyed, black-haired with frontal baldness and, like most engineers, a little introverted. At Google, he led YouTube Live, Gaming, and led engineering for Hangouts. Previously, he won an Oscar and nine feature-length credits for his film design and development software, including
Transformers, Star Trek, the Harry Potter series and Avatar. Prakash, vice president of products and design, is not as big as Petterson, without an Oscar, but he is particularly handsome, with black hair and brown eyes, and he looks straight out of a Hollywood movie set. His youthful appearance does not match twenty years of experience in product development, including the direction of the Google Glass design project. He also worked at Apple for nine years, directly involved in the development of the first iPhone and iPad. This context could, in retrospect, be considered ironic.
Meanwhile, a team of more than twenty engineers and computer scientists at Apple, located just six miles away, aimed to diagnose atrial fibrillation through their watch. They enjoyed Apple's seemingly limitless resources and the company's significant support: the company's general manager, Jeff Williams, head of development and publication of the Apple Watch, had set a Solid vision as an essential medical device for the future. There was no question about the importance and priority of this project when I had the chance to visit Apple as an advisor and review progress. It seemed that their goal would be a goal.
The goal of Apple certainly seemed more achievable at first glance. Determining the level of potassium in the blood might not be something that one would expect with a watch. But, as we will see, the era of deep learning has raised many expectations.
The idea of doing this does not come from AliveCor. At the Mayo Clinic Paul Friedman and his colleagues were busy studying the details of a portion of the ECG known as the T wave and its correlation with blood potassium levels. In medicine, we have known for decades that high T waves can mean high potassium levels and a potassium level higher than 5.0 mEq / L is dangerous. People with kidney disease may develop these potassium levels. The higher the blood level, the greater the risk of sudden death from cardiac arrhythmia, particularly in patients with advanced renal failure or those undergoing hemodialysis. Friedman's conclusions were based on the correlation of ECG and potassium levels in only twelve patients before, during and after dialysis. They published their findings in an obscure journal of cardiac electrophysiology in 2015; the subtitle of the article was "Proof of concept for a new blood test" without blood. "They reported that with variations in potassium level even in the normal range (3.5-5 , 0), differences as low as 0.2 mEq / L could be detected automatically by the ECG, but not by a human examination of the tracing.
Friedman and his team were eager to pursue this idea with the new way of getting electronic ballasts, via smartphones or smartwatches, and integrating AI tools. Instead of contacting major companies such as Medtronic or Apple, they chose to contact AliveCor's general manager, Vic Gundotra, in February 2016, just prior to Petterson and Prakash's membership. Gundotra is another former Google engineer who told me that he had joined AliveCor because he thought that many signals were waiting to be found in an ECG. Finally, at the end of the year, the Mayo Clinic and AliveCor have ratified an agreement to move forward together.
The Mayo Clinic has a remarkable number of patients, which has allowed AliveCor to receive more than 1.3 million ECGs from 12 leads collected from over 20 years of patients, as well as the corresponding blood potassium levels obtained in the three hours following the ECG. to develop an algorithm. But when these data were analyzed, there was a bankruptcy.
Here, the "truths on the ground", the actual potassium (K +) blood levels, are plotted on the x-axis, while the values predicted by the algorithm are on the axis of the there. They are everywhere. An actual value of K + of nearly 7 was estimated at 4.5; the rate of error was unacceptable. After making several trips to Rochester, Minnesota, the AliveCor team worked with this large dataset, many in the depths of winter, in what Gundotra called "three months in the valley of despair." "to try to understand what had happened. false.
Petterson and Prakash and their team dissected the data. At first, they thought it was probably a post mortem autopsy, until they had the idea of a potential return. The Mayo Clinic had filtered its extensive ECG database to provide only ambulatory patients, which skewed the sample towards healthier people and, as one might expect, a fairly limited number of patients. people with high potassium levels. And if all patients hospitalized at the time were analyzed? Not only would this give a higher proportion of people with high levels of potassium, but blood levels would have been closer to the time of the ECG.
They also thought that all the key information might not have been in the T wave, as Friedman's team had thought. So, why not analyze the ECG signal as a whole and negate the human hypothesis that all useful information would have been coded in the T wave? They asked the Mayo Clinic to offer a larger and better dataset. And Mayo went through. Now their algorithm could be tested with 2.8 million ECG incorporating the full ECG pattern instead of just the T wave with 4.28 million potassium levels. And what happened?
Receiver operating characteristics (ROC) curves of true vs false positive rates, with examples of worthless values, good and excellent. Source: Wikipedia (2018)
Eureka! The error rate fell to 1% and the receiver operating characteristic curve (ROC), a measure of the predictive accuracy in which 1.0 is perfect, increased from 0.63 to 0.86 at moment of the scatter plot. We will make a lot of reference to the ROC curves throughout the book, which are considered one of the best ways to show it (highlighting one, and pointing out that the method has been strongly criticized and that efforts are underway to develop better performance indicators). and quantify accuracy – by comparing the rate of true positives to the false positive rate (Figure 4.2). The value indicating the accuracy is the area under the curve, 1.0 being the perfect value, 0.50 being the diagonal line "worthless", the equivalent of a draw at fate. The area of 0.63 initially obtained by AliveCor is considered mediocre. As a rule, 0.80 to 0.90 is considered good, 0.70 to 0.80 fair. They then validated their algorithm prospectively in forty dialysis patients with simultaneous ECG and potassium levels. AliveCor now had the data and algorithm to submit to the FDA for authorization to market the algorithm for detecting high potassium levels on a smart watch.
The AliveCor experiment included essential lessons for anyone wishing to apply AI to medicine. When I asked Petterson what he had learned, he replied, "Do not filter the data too soon …. I was at Google." Vic was at Google. Google.We have already learned this lesson, but you sometimes learn the lesson many times.The machine learning tends to work better if you give it enough data and the rawest data possible. have enough, it should be able to filter the sound by itself. "
"In medicine, you tend not to have enough, they are not search queries, there are not a billion that happen every minute ….. When you have a dataset from one million entries in medicine, it is a giant dataset.And so, the order in which Google works is not a thousand times bigger, but a million times bigger. "Filtering the data so that a person can annotate it manually is a terrible idea. Most AI applications in medicine do not recognize it, but he even said, "It's sort of a seismic shift that, in my opinion, has to affect this industry."
Extract of Deep medicine: how artificial intelligence can make human health again. Copyright © 2019 by Eric Topol. Available from basic books.
[ad_2]
Source link