The Future of medicine
Together with Alpina non-fiction publishing house PostNauka publishes an excerpt from the book "The Future of Medicine. Your health is in your hands" by Eric Topol, doctor, scientist, professor of genomics, director of the Scripps Translational Research Institute, dedicated to modern methods of treatment, information technologies in medicine and new formats of communication between doctor and patient.
The biggest unfulfilled dream in healthcare is the prevention of chronic diseases. In the United States, we spend 80% of the nearly $3 trillion annually allocated to healthcare to cope with the burden of chronic diseases. What if there was a way to stop them at the development stage?
There are other big dreams in medicine.
I still can't forget one chart that I saw in the Economist (Fig. 13.1) more than 20 years ago. In 1994, the magazine predicted that cancer and heart disease would become "curable" by 2040, and the rest of the most serious diseases by 2050. At the same time, life expectancy at birth will increase to 100 years. All this seemed to be a set of too bold expectations, and many of them have not become more real today than they were in 1994. Some prophecies, at least partially, have been realized – for example, robotic surgery and effective treatment of certain types of cystic fibrosis. But, of course, this is not yet "curability". This probably shouldn't be surprising. The word "cure" usually means "restoration of health", or "recovery from illness", or "relief of symptoms of an illness or condition".
Fig. 13.1. Increase in life expectancy at birth and prognosis for the "curability" of most diseases. Source: adapted from “A Survey of the Future of Medicine,” The Economist, March 19, 1994.
There are surprisingly few "cures" in medicine. Some of the examples are the removal of arrhythmias such as atrial fibrillation (in some patients who were lucky), antibiotics in the case of pneumonia, or one of the new types of hepatitis C treatment with recovery in 99% of cases (for the most widespread genotype–1, viral subspecies). Usually, when a person is overtaken by a disease, it must be dealt with somehow. In fact, despite the predictions of The Economist, most scientists who are actively looking for ways to treat cancer hope to turn it into a chronic disease: they have already tempered their ambitions for a cure. When congestive heart failure, chronic obstructive pulmonary disease (COPD), kidney failure, cirrhosis, dementia or serious insufficiency in the work of any organ occurs, there is actually no hope for a cure.
This seems to be a very gloomy forecast. But now, with the arrival of big data, unmanaged algorithms, predictive analytics, machine learning, augmented reality and neuromorphic computing, medicine is being transformed into data science. There is still an opportunity to change medicine for the better, and at least there is a chance of preventing diseases. That is, if there is an accurate signal before the disease has ever manifested in a person - and this information gives good reasons for action – the disease can be prevented.
However, this dream is not just a matter of improving data science. It is indirectly connected with the democratization of medicine. Prospects here are impossible without people mastering the art of observing themselves – remember the double meaning of the term "personalized medicine". Picking up a signal long before any symptoms appear depends on a person's GIS, not on annual doctor visits. With the help of small wireless devices that we carry with us and the Internet of Things, we develop the ability to continuously, very important, monitor our bodies in real time. For a time when such an ability will be fully developed (and eventually it will happen), The Economist's predictions for the next 30 years in medicine do not seem so far-fetched.
The Economist, one might say, overreacted, making such forecasts in 1994. The terms "in-depth data analysis" and "predictive analytics" were definitely not in vogue yet and probably haven't been invented yet. But the concept of using data for predictions, such as actuarial (insurance) statistics in case of life insurance, has been used for a very long time. The difference is that the data sets are now digital, much larger and richer, and they are matched by amazing computing power and algorithmic processing. This is what made it possible for Target to predict the pregnancy of some of its customers, the National Security Agency uses printouts of calls from our phones to identify terrorists, and hospitals predict which patients with congestive heart failure will need hospitalization. And this is what will allow us to "not cut back".
Some things are easy to predict and it is done intuitively. An example is a situation where a public person's illness causes other people to search the Internet for information about this disease or its treatment. You can easily predict that this will happen, and search activity simply reflects the quantitative side of the matter.
And what if you use Google search engines to predict the disease wisely, and not just determine the number of queries? This brings us to the famous Google–related flu story-one of the most cited examples of predictions in healthcare. The Google Flu Trends initiative ("Flu Trends from Google") was launched in 2008 and became known as "a living example of the power of big data analysis." At first, 45 flu-related terms and trends in billions of search queries in 29 countries were tracked. Then the correspondences were derived using unmanaged algorithms to predict the onset of the flu epidemic. By unmanageability, we mean the absence of a given hypothesis – just 50 million search terms and algorithms are doing their job. In widely cited articles in Nature and the Public Library of Science (PLos) One, Google authors (Figure 13.2) claimed their ability to use Internet search logs to create daily estimates of flu infection, unlike conventional methods that provide a time lag of one to two weeks. And further, in 2011: "The Google Flu Trends initiative can provide timely and accurate estimates of the incidence of influenza in the United States, especially during the peak of the epidemic, even in the case of a new form of influenza."
Figure 13.2. Data from sanitary and epidemiological centers (light line) in comparison with Google Flu Trends data (dark line) on flu outbreaks in the mid-Atlantic states of the USA. Source: J. Ginsberg et al., “Detecting Influenza Epidemic Using Search Engine Query Data,” Nature 457 (2009): 1012-1015. Reprinted with permission.
But the beginning of 2013 was accompanied by a storm of contradictions: it turned out that Google Flu Trends greatly overestimated the flu outbreak (Fig. 13.3).
Later, a group of four highly respected data scientists wrote in Science that Google Flu Trends systematically overestimated the spread of influenza every week since August 2011. The group went on to criticize the "arrogance of big data", "the widespread perception that big data replaces rather than complements traditional data collection and analysis.
Fig. 13.3. Google Flu Trends has overestimated the flu. Flu Near You ("Flu near You") is another initiative launched in 2011. Source: D. Butler, “When Google Got Flu Wrong," Nature 494 (2013):155-156. Reprinted with permission.
They scolded the "dynamics of the algorithm" of Google Flu Trends (GFT), pointing out that 45 terms used in search queries were not documented, key elements, such as the main search terms, were not presented in publications, and the original algorithm was not subjected to constant adjustments and rechecking. Moreover, although the GFT algorithm was static, the search engine itself was constantly changing, having undergone no less than 600 revisions per year, which was not taken into account. Many other editorial writers have also spoken out on this issue. Most of them paid attention to the relationship instead of cause-and-effect relationships and the critical lack of context. Sampling methods were also criticized, since crowdsourcing was limited to those who performed Google searches.
In addition, there was a serious analytical problem: GFT conducted so many data comparisons that there was a chance of getting random results. All of these can be seen as common traps when we try to understand the world through data. As Krenchel and Madsbjerg wrote in Wired: "The arrogance of big data is not that we are too confident in a set of algorithms and methods that are not yet in general. Rather, the problem is the blind belief that it is enough to grind numbers while sitting at a computer to understand the world around us in its entirety." We need answers, not just data. Tim Harford put it bluntly in the Financial Times: "Big data is already here, but there are no great insights."
Some began to defend the GFT, pointing out that the data was just an addition to the sanitary and epidemiological centers, and Google never claimed to have a magic tool. The most balanced point of view was expressed by Gary Marcus and Ernest Davis in their article "Eight (no, nine!) problems with Big data" (Eight (No, Nine!) Problems With Big Data). I have already addressed many of their conclusions, but the opinion of Marcus and Davis about the shameless advertising of big data and about what big data can (and cannot) deserves special mention: "Big data is everywhere. It seems that everyone collects them, analyzes them, makes money on it and glorifies their power or is afraid of them ... Big data is not going anywhere, as it should be. But let's be realistic: this is an important resource for anyone who analyzes data, not a silver bullet." Despite the problems with GFT, such steps do not lead anywhere. An alternative and more recent approach is to predict the outbreak using a smaller base of people who actively kept in touch on Twitter – the so–called "central nodes", when people essentially act as sensors. This made it possible to detect outbreaks of viral diseases seven days faster than when the population as a whole was considered.
Similarly, the HealthMap algorithm, which searches tens of thousands of social networks and news media, was able to predict the outbreak of Ebola in 2014 in West Africa nine days ahead of the World Health Organization. I delved into the history associated with Google and flu and outbreaks of infectious diseases, because they show the early stages of the path we are on and show how we can get lost using large amounts of data for predictions in medicine. But knowing how we lost our way is important if we are going to move along it.
Predictions at the individual level
Compared to data for the entire population, as in the case of Google Flu Trends, a more powerful effect is achieved by combining detailed data of an individual with detailed data of the rest of the population. You've come across this before. For example, Pandora has a database of preferred songs for more than 200 million registered users who have clicked on the "like" or "dislike" buttons over 35 million times in total. The company knows who listens to music when driving, who has an Android and who has an iPhone and where each of them lives. As a result, it is possible to predict not only what kind of music the listener will like, but even his political preferences, and the company has already used this in targeted political advertising during the presidential election campaign and congressional elections. Eric Bishke, Pandora's chief scientist, believes that their data collection programs allow you to get to the very essence of their users. And this is true, because to get to the bottom of it, they integrate two layers of big data – your data and the data of millions of other people.
Using data trading companies such as Acxiom (which were discussed in the previous chapter), the University of Pittsburgh Medical Center conducts in-depth analysis of its patients' data, including characteristic behavior during shopping, to predict the likelihood of using first aid services. The Health Organization of North and South Carolina does the same, collecting data on credit cards of customers – 2 million people in their region to identify patients with a high risk of diseases (for example, through purchases of fast food, cigarettes, alcoholic beverages and medicines). The predictive model used in Pittsburgh showed that consumers who make the most purchases online and order goods by mail are more likely to turn to first aid points, which health organizations do not welcome at all. The discovered relationships acquire new details over time, when information about current patients is received repeatedly and more patients are included in the system in order to better predict certain processes. But questions of confidentiality and ethics remain.
These examples can be seen as a rudimentary form of artificial intelligence–machines or software demonstrating human-like intelligence. Other examples that may already surround you include personal digital assistants like Google Now, Future Control, Cortana and SwiftKey, which combine information from emails, SMS, diaries, notebooks, search query history, locations, purchases, who you spend time with, your preferences in art and your behavior in the past. Based on what they learn from this information, these apps appear on your screen to remind you of an upcoming meeting, show traffic jams on your route, or report news about your flight. Reading what your friends write on Twitter, Future Control, they can give you advice: "Your girlfriend is sad, send her flowers." SwiftKey even calculates your typing errors and corrects them if you press the wrong key all the time. Google Now works with airlines and event organizers to have access to ticket information, and can even listen to the sound of your TV to provide you with a TV program in advance. As you can guess, these are much more powerful features than the matching search that drives Google Flu Trends, and they are directly related to medicine.
Such predictive power relies solely on machine learning, a key feature of artificial intelligence. The more data is entered into a program or computer, the more they learn, the better the algorithms and, presumably, the smarter they become.
Machine learning and artificial intelligence techniques are what ensured the triumph of the IBM Watson supercomputer over people in the Jeopardy TV quiz! (Take a chance!). It was necessary to quickly answer difficult questions, the answers to which could not be found using the Google search engine. IBM Watson were trained to answer hundreds of thousands of questions that were asked in previous Jeopardy! quiz games, armed with all the information from Wikipedia and programmed for predictive modeling. This is not a prediction of the future, but simply a prediction that IBM Watson has the right answer. At the heart of the supercomputer's predictive capabilities was an impressive portfolio of systems for machine learning, including Bayes networks, Markov chains, the support vector machine method and genetic algorithms. I won't go into it any further: I'm not smart enough to understand all this, and, fortunately, it doesn't really apply to where we are going now. Another subspecies of artificial intelligence and machine learning, known as deep learning, is of great importance for medicine.
Deep learning is behind Siri's ability to decode speech, as well as Google Brain's experiments with pattern recognition. Researchers from Google X extracted 10 million images from YouTube videos and launched them into a network of 1,000 computers to see what Google Brain, which has a million simulated neurons and a billion simulated synapses, is able to offer on its own. The answer is cats. The Internet, at least the YouTube segment (which occupies a very significant part of it), is full of videos of cats. In addition to cat identification, this discovery illustrated cognitive computing, also known as neuromorphic. If computers can compete with the human brain, as the theory says, then it is possible to achieve the transition of their functional capabilities in terms of perception, action and understanding to the next level. Progress in neuromorphic computing is going at breakneck speed. Last year, the accuracy of computer vision – for example, recognition of pedestrians, helmets, cyclists, cars – improved from 23% to 44%, while the error rate decreased from 12% to less than 7%.
Despite the achievements of Google Brain, we still have nothing to brag about. The human brain works at low power, about 20 watts, and a supercomputer requires millions of watts to work. While the brain does not need to be programmed (even if it sometimes seems that it is programmed) and it loses neurons throughout its life without significant functional exhaustion, a computer that has lost a single chip can break down, and usually machines cannot adapt to the world they interact with. Gary Marcus, a neuroscientist at New York University, put this neuromorphic task in perspective: "At times like these, I find it useful to recall a basic truth: the human brain is the most complex organ in the universe, and we still have no idea how it works. Who said that copying its amazing power would be easy?" Nevertheless, there has been quite a lot of progress in speech recognition, faces, gestures and pictures, in which the human brain is so strong and the computer is weak. I have attended many conferences and lectured in different countries, with simultaneous translation, and I was particularly struck by one achievement: Richard Rashid, who once headed the scientific department at Microsoft, gave a lecture in China, and the computer not only synchronously gave it in hieroglyphs, but also translated it into Chinese with the (simulated) voice of Rashid himself. Facebook's DeepFace program, with the world's largest photo library, can determine whether two photos belong to the same person with 97.25% accuracy. The consequences for medicine are obvious.
Scientists are already showing that computers are able to recognize facial expressions, such as pain, more accurately than humans, and there is amazing progress in facial recognition by computers. Computer scientists from Stanford University used a cluster of 1,600 computers to prepare for image recognition, training was conducted on 20,000 different objects. More relevant to our topic is that they used deep learning tools to determine whether a sample taken during a biopsy in the case of breast cancer is malignant. Andrew Beck of Harvard University has developed a computerized system for diagnosing breast cancer and predicting the chances of survival based on automatic image processing. It turned out that training based on computer data processing provides greater accuracy in comparison with pathologists, and this helped to recognize new features that remained unnoticed for many years. And we should not forget about the active support for the development of artificial intelligence, which made it possible to create seeing and hearing devices. The Orcam sensor camera is installed on the glasses of visually impaired people, it sees objects and transmits this information through an earphone using bone conduction. The GN ReSound Linx and Starkey hearing aids are smartphone–connected applications that "provide people who have lost their hearing with the opportunity to hear better than those who hear normally."
There are wheelchairs for people without four limbs, controlled by thought, in the spirit of the bionic future. Therefore, the ability of artificial intelligence to transform the material world in medicine should not be overlooked. Of course, technology can easily connect with robotics. At the University of California, San Francisco, the hospital pharmacy is fully automated, and the robotic delivery of medicines is still going on without a single error.
Portal "Eternal youth" http://vechnayamolodost.ru