Artificial intelligence

In this section we will learn what artificial intelligence is, what it can do, how its created and how it can affect healthcare.

This two-minute video will introduce you to the basics of artificial intelligence

What is artificial intelligence?

AI is a set of instructions which are written in a computer program (software). The instructions run a computer which performs mathematical tests on data. Which mathematical tests are used and what the result looks like depends on what problem we wanted the AI to solve. There are lots of different types of AI. The instructions that allow the AI to work are called an ‘algorithm’.

AI can be complex or simple. Most of us use simple AI every day. When we search the internet the search returns web pages which may be relevant. AI is what puts all those web pages into the order you see on your screen. Banks use AI to help show when bank details have been stolen. Online TV and film companies like Netflix and Amazon use AI to make suggestions for other programmes you might want to watch based on what you have watched before.

This podcast is a conversation between Jonathan Gregory, and Danny Ruta. They are both doctors who work with AI experts to create and test AI in healthcare. Here they talk about AI and how it is created. 

What is an algorithm?

An algorithm is a set of clear instructions used to solve a problem. A recipe we follow to bake a cake solves the problem of ‘I don’t have a cake’. If we take the ingredients and follow the recipe, we should get a cake. An algorithm is a recipe, a set of rules, that only involve maths. By following this maths recipe, the algorithm solves problems.

Is AI the same as the human brain?

Some people are immediately worried or concerned when we talk about AI. They imagine human like robots with super intelligence. These super AI robots do not exist and are not going to exist any time soon.

One of the most amazing things about humans is that we are so good at so many different things, like language, mathematics or using common sense when things don’t go to plan. The list of our human abilities is enormous. There is not a single AI system that can do all of the things that a human can do. 

Most scientists think we are many years away from having an AI that can match humans across all the skills we have. However, there are AI systems that are ‘better' than humans at some specific tasks. These are narrow highly focused tasks. These AI systems have what is called narrow artificial intelligence. Sometimes they are better than humans just because they can work faster. Sometimes they are better because they can see patterns in data that humans can’t because the data is too large or complicated.

But these narrow AI systems are only good at one thing. A narrow AI that can spot a cancer on an x-ray would not be able to tell if a blood result was normal. In the world around us today we only have narrow artificial intelligence and this is the type of artificial intelligence that is starting to be used in healthcare today. 

Who makes the algorithm (recipe) that the computer follows?

A computer follows a program that allows it to solve problems using mathematics. That program contains an algorithm, and it is the algorithm that gives the computer ‘artificial intelligence’. 

Algorithms made by humans

The algorithm can be made entirely by humans. The computer is given a fixed set of rules that it follows. 

For example, if we make some software that contains the English dictionary and we put that software into a computer, we can then search on the computer for the definition of words.  The computer is just giving us the definition we gave it but it is doing it much more quickly than we can by looking it up in a book. 

In the same way we can give the computer all the ‘rules ’for treating breast cancer. We write software with all the information on what treatments are used for different sizes and types of breast cancer. We can then give the AI details of a patient’s cancer and ask what treatments the guidelines recommend. 

The AI would not be ‘deciding ’what treatment to give. It would be very quickly checking the patient’s clinical data against the guidelines and rules that humans had created based on research.

Medical science has become so advanced that it is harder and harder for doctors to keep up to date with every new treatment so an AI like this can be helpful. 

These are rule-based algorithms, sometimes called expert systems. This technology has been available for years. This is ‘old fashioned’ AI. The computer algorithms must be programmed into the computer by humans. Rule-based algorithms can be very useful for making simple tasks happen automatically. But they are limited in the types of problems they can solve because humans have to make the recipe (algorithm) for the computer.

Example of a Basic Algorithm:

At their most simple, rule-based AI involves the computer following basic instructions. These instructions can be described as IF and THEN.

For example, a simple AI which had the task of highlighting abnormal blood test results might follow rules like this:

IF result bigger than 10 THEN highlight in red.

That way humans can just check any numbers that are highlighted in red and might only skim the other results.

The AI is not creating new suggestions it is only doing a job we told it to do. A rule-based AI like the one we talked about in the text that could help doctors decide what treatments to give patients with breast cancer would be much more complicated. However, at its heart it would still be following lots of instructions. 

For example, IF the age is greater than 75 THEN possible treatments are A, B, or C. 

But the complexity would be much greater as there would be lots of IF-THEN outputs which the AI would need to combine.

Algorithms made by computers – machine learning

The computer is given lots of data (humans have to be very careful to only use good quality fair data). The computer looks for patterns in the data. Gradually it turns these patterns into mathematical formulas that allow it to solve the problem we set.  The AI will adjust the mathematical formulas to get better at solving the problem. Different algorithms are used for different types of problems. For example trying to find patterns in data, identifying objects in an image or making predictions from data. Humans give set up the AI with the right basic type of algorithm and then through machine learning the algorithm changes to hopefully get better at solving the problem it has been given.

How does this work in real life and could it work in healthcare?

Let’s imagine we are trying to find new ways to detect cancer at its early stages from blood tests. If we made a rule-based algorithm we would need to tell the computer what was normal and abnormal, which results might be linked to cancer. This algorithm might help to review thousands of blood test results more quickly than humans could so it might be helpful. But it would be spotting things that we could spot if we only had the time to look carefully. It would not discover new clues in the data. 

If we used machine learning, we would give the computer blood test results, including people whom we knew had cancer and people who didn’t. The machine learning algorithm could be asked to predict which people in the data had cancer and which didn’t. Humans would check the result and give feedback where the AI was wrong. The AI would then adjust the algorithm to try and be more accurate and humans would check it again.

This would be repeated until the AI was really good at predicting which people had cancer in the data we had used to train it. We would then give the computer data on a new set of patients and see if it still worked well. We would hope that in the end, the AI could alert us if a blood test result showed warning signs of cancer. But the AI would have done this in a different way than humans would have thought of. It would almost certainly see patterns in the data that humans could not see and use these patterns to make predictions.

Are computers really learning?

When we start talking about computers ‘learning’, is when some people start to get worried; scary science fiction movies spring to mind. Machines or computers don’t learn in the same way that humans learn. 

What machines or computers are doing is using trial and error to get better and improve their performance. They can do thousands of mathematical calculations per minute, which helps them improve quickly. This is what is meant by ‘learning’, the computers don’t understand the data. They are finding the quickest or most correct mathematical tests to use on the data to get the results the human has asked for.

With rule-based systems, when humans make the recipe (the algorithm) it stays the same unless humans change it. With machine learning the computer changes the algorithms and tests it thousands of times. The computer is trying to find the best algorithm to give the correct answers on the data it has been given.

Some AI looks like it has human intelligence

There have been recent advances in a type of AI called generative AI. Chat GPT and Google Bard are examples of generative AI. This type of AI uses machine learning. It is still narrow AI but it uses lots of algorithms added together to perform tasks. This type of AI is used to create (generate) things for example text, pictures or audio. It is very powerful, but it is still not thinking like humans. It is using probability to give the most likely answer to a question. It has been trained on over 300 billion words! 

However, that does not mean that it is definitely free from bias and it can give misleading results to some tasks. 

We now know that AI is built using algorithms, but what exactly are they? Watch this two-minute video to find out.

Real World Examples of Rules Based and Machine Learning Narrow AI:

Go is an ancient Chinese board game. The number of possible moves is even higher than the number of moves in a game of chess. It has been calculated that there are more possible moves in a game of Go than the number of atoms in our universe!

A company called Deep Mind developed a machine learning AI called AlphaGo which beat the human Go World Champion, Lee Sedol. The scientists that programmed the computer did not teach the computer how to play, they just gave it the rules of the game. The algorithm then evolved by playing thousands of matches against amateurs and then professional players. The computer analysed what combinations of moves led to a win and which lost. It gradually improved until the AI was good enough to beat the world champion.

Narrow AI Can Only Solve a Narrow Set of Problems.

The example of an AI algorithm being better than humans at a complex game like Go can raise concerns about the power of some of these AI systems. But remember, these narrow AIs are only good at the one thing.

If we gave the AlphaGo AI an IQ test it would fail! It would not be able to answer any of the questions and would have an IQ of zero!

We could build an AI to answer the IQ test and over time we could probably get it to pass the test with the highest possible mark – an IQ of 201. This could look impressive, but the AI would not really have a high IQ. If we tried to use this AI for anything else, it would be useless.

Now we understand what AI is and how it is created, we will learn about the problems that can happen when we use AI in healthcare. Most

At the moment, AI is being regulated and approved in the same way medical equipment is approved. That is OK for rule-based AI. 

Rule-based AI doesn’t change unless humans deliberately change it, and they would know and understand those changes. But this doesn’t work in the same way for machine learning AI. As the machine learning AI is shown new data it will change its algorithm. Usually once a machine learning algorithm is being used in the real world it must be retrained by humans who will give it new sets of data based on where the algorithm is being used. 

How is AI regulated in the UK?

Most AI in healthcare will be classed as a medical device. A medical device is anything, that is not a medicine, that can be used for a medical purpose such as making a diagnosis or treating someone. It includes things such as hip replacements and wound dressings but also software. For example, software in an app that helps people with diabetes to manage their dose of insulin.

The market for any technology used in the UK which is a medical device is regulated by the Medicine and Healthcare products Regulatory Agency (MHRA).

AI which is used to help diagnose, treat or prevent health problems would be classified as a medical device. This is regulated and approved in the same way other medical devices are approved.

Rule-based AI doesn’t change unless humans deliberately change it, and they would know and understand those changes. That makes it straightforward to tell the regulator what changes have been made, how these might affect safety and the regulator can decide if they approve the new rule-based AI for use. As humans need decide what changes to make in rule based AI the rate of change will be slow.

However, as you have learnt machine learning AI adjusts the mathematical formulas it uses as it is given new data. The initial device approval may only relate to one version of the machine learning AI. This helps ensure safety but it means that the way changes to the algorithms that could improve the AI (and this could be quite often early in the use of an AI) will need to be approved as well.  This is the challenge we face and we are trying to find a balance between allowing AI to develop quickly so it can help us make healthcare better whilst being safe and being sure no harm is caused. 

We have now covered some of the problems that we have to overcome if we are going to use AI in healthcare. Some of these problems are difficult to solve. You might ask, ‘why bother?’, ‘Why do we need AI in healthcare?’ 

We will look at some possible benefits and then you can decide for yourself whether you think AI in healthcare is a good thing or not. It’s not all bad news! AI can correct human mistakes.

Example: Zooming in on knee pain

Knee pain is common.  One cause of knee pain is arthritis of the knee. This is where the lining of the knee joint wears out and becomes painful. Surgeons use x-rays to help diagnose arthritis and decide if a patient would be helped by knee replacement surgery. 

Diagnosing arthritis on x-rays is a skill that doctors develop during their training. The arthritis often develops in a standard pattern that is easy for experienced doctors to find. Not everyone who has pain in the knee has arthritis. Sometimes no cause for the pain is found and people must live with long term knee pain. Researchers noticed that more patients who were black were diagnosed with long term knee pain. These patients had x-rays which the surgeons did not think showed arthritis. 

They also noticed that fewer people who were black had knee replacements, so they tried to find out why. The data about these patients was reviewed using AI. The scientists found that the surgeons were right about some of these patients and wrong about others. Some of these patients did have arthritis when the surgeons thought they didn’t, but it was not the normal pattern of arthritis that surgeons had been taught to spot.

Because surgeons had not realised you could get this different pattern of arthritis, they didn’t know to look for it. This meant some patients had their diagnosis of arthritis missed and so they didn’t get a knee replacement. 

We needed something to look at the data from a new perspective.  The AI wasn’t studying the data in the way humans would and this meant it could find something we had been missing. This shows us that whilst there can be problems with AI there are also problems with humans. We make mistakes and we don’t always realise why we are making them. AI gives us a chance to notice things in the world around us that we haven’t yet seen.  There is a risk of bias with AI if the data is biased but there is also a hope that AI can challenge some of our human bias and make healthcare fairer for all patients.

Examples of how AI might improve healthcare:

• There are not enough x-ray specialist doctors, so patients wait a long time for scan results
  AI might be able to check all the x-rays and scans and find those that are normal. It could create an automatic result that was sent to the patient and their doctor the same day as the test was done. Doctors could use their expertise looking at abnormal x-rays, or scans that the AI was uncertain about.
• Doctors sometimes miss an abnormality on a scan or blood test
  AI might be able to double check the doctors. By reviewing all the tests, it could alert humans to double check if the AI thinks something is abnormal, but the human doctor hadn’t noticed.
• Doctors spend a lot of time checking blood test results which are normal
  An AI could automatically tell patients they had a normal blood test result and file the result in the electronic medical records or book an appointment for clinic.
• Nurses and healthcare assistants spend a lot of time checking hospital patients blood pressure and temperature
  AI can assess the results from devices that automatically measure temperature and blood pressure. The AI could alert the nurses to problems but if everything is ok the nurses can carry on with all the other work they have to do.
  
• Booking appointments for clinics and scans takes a lot of time and effort for human staff
  AI will be able to automatically schedule clinics and scan appointments and try and help them be booked so they run on time by not being over or under booked.
  

How might AI free humans to be more human?

Computers have made a big difference to healthcare. Many of these changes have been for the better but computers can cause problems. Before computers were commonly used, an appointment with a doctor would involve the doctor asking questions and listening to the answers. The doctor would also be trying to build a relationship with the patient. They would look at the patient’s body language to help understand how a patient was feeling. One of the reasons people complained about doctors’ handwriting is that they would write very short notes very quickly at the end of the appointment so they could listen carefully to the patient when they were speaking.

Once computers came along the doctors started to spend more time looking at the computer and typing things into it than looking at the patients. This has meant that some patients do not feel like the doctor has listened to them properly. It has also caused some doctors to not enjoy their work as they want to be ‘with ’their patients, not looking at a computer.

AI might be able to solve this. It might help us to bring the benefits that computers offer but reduce the negative impact they can have on doctor and patient communication. As long ago as 2006, software was developed which could take what the doctor was saying during a conversation with a patient and start to fill in the paperwork. For example, if the doctor was going to refer the patient, the computer would start to complete the referral letter from what the doctor had said. This reduced the paperwork doctors had to fill in after each appointment.

Unfortunately, the methods used in 2006 could not work in every situation so it did not become widespread in the NHS. Now we have a better understanding of AI, researchers are again trying to solve this problem. Some very large companies are trying to make AI which will mean that doctors won’t have to type into the keyboard of a computer. They are trying to make an AI system that will pick words out of the conversation and bring up test results automatically so the doctor can discuss them. It will start to automatically order tests the doctor thinks are needed and write the letter that is sent to the patient and their GP after every appointment.

Computers were brought into the NHS to solve some problems but they created others. AI might help us to have the benefits of computers but stop them getting in the way of patients and doctors having proper conversations.

The use of new technology in healthcare is always difficult and problems will occur. Whenever there has been a brand-new operation, medical device, or drug treatment there have often been problems to overcome. In just the same way, the use of AI in healthcare will cause some problems.

In Conclusion

We think that none of the possible problems on their own are a reason not to use AI in healthcare. Being a doctor, nurse, midwife, pharmacist or physiotherapist is harder now than at any other time due to the huge advances in medical science. There is now too much for any human to know or to keep up to date with. AI can help people do this. It can release professionals from administration tasks giving them more time to spend with patients and supporting people to stay healthy or to get better.

AI can study data in new ways and help humans to better understand health and disease. We think that everyone needs to be careful with AI. We need to work hard to make it safe by understanding how it might go wrong and how to prevent this from happening. AI offers a real opportunity to deliver better care for more people around the world. A careful approach will allow us to receive help from what AI has to offer while minimising the risk of harm.

Resources

To learn more visit our Resources page.

This two-minute video will introduce you to the basics of artificial intelligence

This podcast is a conversation between Jonathan Gregory, and Danny Ruta, about using AI in healthcare. Jonathan and Danny are both doctors who work with AI experts to create and test AI in healthcare.

We now know that AI is built using algorithms, but what exactly are they? Watch this two-minute video to find out.

This podcast is a conversation between Jonathan Gregory, and Danny Ruta. They are both doctors who work with AI experts to create and test AI in healthcare. Here they talk about AI and how it is created.