How AI may solve our health problems

The development of artificial intelligence is moving quickly and will change our life – not least when we go to the doctor. We are taking the temperature of the AI fever within medicine.

illustration av AI i vården.
Illustrations: Björn Öberg, first published in the magazine Medicinsk Vetenskap in 2017.

”It is like a baby who learns something new every day”, says Max Gordon, chief physician and researcher in orthopaedics at the Department of Clinical Sciences, Danderyd Hospital.  

He is talking about an artificial intelligence – a computer program with the human brain as a prototype – which is currently learning to interpret X-ray pictures in his lab. Instead of a doctor identifying fractures and assessing which patients require surgery, the idea is that computers should be able to do this in the future. There would be several benefits of such an artificial helper in hospitals. 

“An artificial intelligence never gets tired and can work unsocial hours.  Doctors will instead get the time to look more closely at the complicated cases and meet patients,” says Max Gordon.

Instead of regular DNA the AI baby has algorithms, mathematical instructions which it received from its programmer. The program is not designed to do everything right, like a regular computer program, but to be able to learn a task through practise. 

Max Gordon, credit: Stefan Zimmerman.

“That is the intelligent part. It is similar to the way the human brain learns by processing information in a network of cells,” says Max Gordon.  Nobody explains to the program in Max Gordon’s lab how to behave in order to interpret the photos, instead it learns this on its own, he clarifies.

However, at the start during the practise period, Max Gordon needs to be available in the background to give the lead, adjust the algorithms slightly and confirm when the program gives the right answer. His research results, which have not been published yet, show that it is currently right for over 90 per cent of the cases, which is in line with a regular radiologist. 

The concept of creating machines which can think has existed since Antiquity and eventually resulted in the development of the first computers. However, many people thought that it was impossible to teach a computer to think like a human being. Nonetheless, the AI advocates were persistent. In 1948 the mathematical genius John von Neumann said: “If you will tell me precisely what it is that a machine cannot do, then I can always make a machine which will do just that”.

“The theories behind artificial intelligence are old but they have been difficult to implement in practise. Just a few years ago computing power reached a level which actually made the programs start to function,” says Max Gordon, who has worked with programing since his younger years. 

Classify millions of photos

The breakthrought occured in 2012 in a competition where computer programs were assigned to classify millions of photos retrieved online. In one corner of the ring there were traditional programs, specifically written for the purpose by experts in computer vision. In the other corner of the ring there were AI programs with no previous knowledge, but with the potential to learn.

The competition was won unexpectedly by a significant margin by the AI program AlexNet, created by the researcher Alex Krizhevsky at the University of Toronto, after 5 days of practise. Three years later an AI program won again – but now at a level which not only surpassed other computer programs, but human beings as well.

Human beings have also been defeated by computers in games such as chess, the quiz Jeopardy and board game Go. Thus it seems as though AI programs are about to move on from the baby phase and reach a superhuman level in several areas. The successful algorithms are immediately adopted by several developers who customise them to new tasks. For example, the algorithm for AlexNet was downloaded by Max Gordon who started to train it to interpret X-ray pictures. 

Portrait of Mattias Nilsson Benfatto.
Mattias Nilsson Benfatto, credit: Stefan Zimmerman.

Another promising algorithm was further developed by Mattias Nilsson Benfatto, computational linguist and researcher at the Department of Clinical Neuroscience, Karolinska Institutet. He trained it for a completely different task, namely to detect dyslexia among school pupils.

“Dyslexia is not located in the eyes, but is the result of the brain being unable to decode text. By measuring how the eyes move, we can get a view of the linguistic processes that go on inside the head when a person reads or writes,” he says.

One way of identifying children who are in the risk zone of dyslexia, so that they can receive assistance and support, is to let entire school classes take language tests. Another way is to measure eye movements.

“A strength of measuring eye movements is that it becomes more objective. In traditional paper tests there are several sources of error, as pupils, for example, need to handle papers and a pencil, and the tests need to be marked by someone. It is also faster, the measurement only takes a few minutes,” says Mattias Nilsson Benfatto.

 Very large amounts of data

The problem is only that the measurements result in very large amounts of data which is difficult to analyse manually. By training the AI algorithm, Mattias Nilsson Benfatto and his colleague Gustaf Öqvist Seimyr have now automated the analysis, which has made eye movement measurements a more useful tool. In a research project, the results of which are currently being analysed,

Mattias Nilsson Benfatto and his colleagues have conducted eye movement measurements in school classes in Järfälla and Trosa. They want to find out about the reliability of the automatic analysis and at what age dyslexia can be identified at the earliest. 

“Our results so far show a high level of conformity, 90 per cent, between the AI program’s assessment and traditional assessments,” says Mattias Nilsson Benfatto. 

Portrait of Mikael Huss.
Mikael Huss, private image.

This is good enough for Mattias Nilsson Benfatto and Gustaf Öqvist Seimyr to now use the technology within the framework of their company, Optolexia, which offers dyslexia scanning to schools. The technology which Max Gordon and Mattias Nilsson Benfatto use is referred to as ‘supervised learning,’ as it requires a human being to supervise the learning with solutions being available.   

Another, more freethinking artificial intelligence, can be found in a computer at SciLifeLab in Solna.

 “Artificial intelligence may be a good way of automating things which humans can also do. But many hope that artificial intelligence will also be able to contribute with something completely new, perhaps detect things which no human being has thought about looking for,” says  Mikael Huss,  researcher in bioinformatics at SciLifeLab.

Super computer

As an example he mentions the super computer Deepmind, which beat one of the world’s best players in the board game Go in 2016. First the computer observed human matches, then it played matches against itself and learnt from its mistakes. The result was a playing style which is not human. 

“It played in a way which, according to what the Go players have learnt from childhood, ought to be completely wrong. So it had seen something, found a pattern, which humans had completely missed,” says Mikael Huss.

If such a system was applied within medicine, it would find new scientific links, for example, previously unknown links between different diseases or, well... something completely unexpected. This type of breakthrough has not occurred yet, but researchers are working on it. Mikael Huss has started little by little by allowing an AI program to analyse data open-mindedly on the content in some cells, without explaining what is right and wrong.

The information was obtained from two studies which other researchers conducted with what is referred to as ‘single cell analysis’. The aim was to see whether the program could find a way to screen the information on its own and create a compact description, a type of definition, of what characterises different cell types.   

“It worked quite well, the program made roughly the same assessment as the researchers,” he says.

The experiment was conducted as part of a degree project which Mikael Huss supervised. Even though the results were promising, the reliability of results was unclear. The program needs to become much more proficient in its task in order to reach a level which would make it workable practically. Reaching this level necessitates practise on many more examples first. In this context Mikael Huss sees a challenge for artificial intelligence within the medical area.

“An AI program finds it easiest to learn if it can practise on lots of examples, each of which can be described with relatively few parameters.  But within medicine often the opposite applies. Researchers often have problems collecting information on an adequate number of patients for their studies, but each case contains a large amount of data which can be interpreted in different ways,” he says.  

Thrives like fish in water

However, in the Internet’s sea of information AI thrives like fish in water. Millions of practice examples are currently being input there in the AI program, in the form of news articles and posts on social media which are translated automatically. You have probably seen them in your Facebook flow.

The computer company IBM is trying to give its computer system Watson similar conditions, but for the scientific area by inundating it with all text-based scientific information the company comes across, in the form of scientific articles, textbooks and journals. The company has also paid billions to purchase 30 billion X-ray pictures which Watson can practise on.  

Illustration of AI.

According to IBM, 80 per cent of all health data has been unavailable until now as it is unstructured. However, Watson will be able to see the information and has a dizzying reading speed – almost 200 million pages of text in three seconds. Marketing texts portray superb plans for “Dr Watson” which will initially learn everything about cancer research. The computer system has apparently already made several cancer diagnoses where human doctors have failed. Max Gordon says that there is great enthusiasm in Sweden as well, but believes that it will take a while for artificial intelligence to assume a significant role within medical care. At least in Sweden.

“I’ve spent a lot of my research time applying for different permits and trying to understand the rules which apply. Everyone is enthusiastic but there are not any good established paths, as a result of which the preparations in Sweden take longer than they do, for example, for research groups in the US which I have been in touch with,” he says.  

A poor translation will not result in anybody’s death, which is why artificial intelligence can be rampant in risk-free environments, such as in your Facebook flow. But an artificial intelligence which will work in medical care must be a fully-fledged pro from day one and consequently it may take a while before it is assigned important duties.

A matter of time

However, Max Gordon believes that it is mainly a matter of time – artificial intelligence will change medical care. This does not necessarily mean that artificial intelligence will replace medical care personnel, but it will lead to a division of tasks between things which humans and machines respectively are best at.  

“I’m convinced that diagnostics will mainly take place with artificial intelligence. I can’t see any reason for why that won’t be the case, as computers will soon be able to do it in a faster, better and safer manner than human beings,” he says.

But communicating a diagnosis to the patient with empathetic sensitivity, or making the right choice between two treatments with their subjective advantages and disadvantages – such things are still done best by humans.  

“Computers can make risk assessments and present different options. But they don’t have any understanding of the subjective experiences of humans. I think that the role of doctors will progress and focus more on the meeting with patients,” predicts Max Gordon.

Solve limited tasks

The artificial intelligence which has been developed until now can solve limited tasks.  Whether this now means that machines can think is uncertain. If that is the case then the thinking is very narrow and single-minded, compared to what us humans do when we think. But maybe at one point what is referred to as general artificial intelligence will develop – a computer which can think freely and learn anything.   

Some thinkers, such as the Microsoft founder Bill Gates, entrepreneur and future visionary Elon Musk and scientist Stephen Hawkings, warn that the development of artificial intelligence may be the biggest mistake of mankind. What is there to guarantee that a super intelligence, which is superior to human beings in all areas, will be positive towards us?   

Others like Ray Kurzweil, Director of Engineering at Google, believe that human beings and computers will merge together to a new super intelligent life form, with the ability to solve all problems in the world. 

“I don’t think about things like that, it seems very far away from where we are right now” says Mikael Huss, even though he does not object to the reasoning itself. 

Mikael Huss agrees that major changes are underway, but does not believe that we will get to see this artificial intelligence taking over the world.

“Artificial intelligence is both overhyped and underhyped. A lot of things in society will be managed by artificial intelligence in the future, I really believe that. Step by step we will accept things which facilitate our life, not least within areas such as research and medical care. But when that happens we will already be used to it, just like with all technology, it won’t feel like science fiction anymore,” he says. 

Text: Ola Danielsson, first published in Swedish in Medicinsk Vetenskap no 1/2017

Facts: 3 famous AI:s to take into account

Eliza is a basic chat robot developed in 1966. Despite the fact that Eliza was completely unintelligent, many felt that she really showed understanding of psychological problems, among other things. Eliza’s standard phrases were “how are you doing now?” and “tell me more”.

AlphaGo is a computer program that plays the strategically advanced board game Go. When AlphaGo defeated the champion Lee Sedol in February 2016, the journal Science considered, as one of nine options, choosing it as Breakthrough of the Year (instead the award went to the discovery of gravitational waves).

Watson is a computer system created by IBM. Originally developed to answer questions on the quiz show Jeopardy!, it is currently pursuing a medical career.  Its duties will include making cancer diagnoses and helping out where there is a shortage of doctors in India.

Content reviewer: