Unlike their more modern large language model counterparts, artificial neural networks require human input to learn and function.
ANNs have been around since the 1950s. They started taking hold in healthcare in the 1990s. Within the last 10 to 20 years, they’ve been having an impact on healthcare. Today, ANNs have taken a step up – many are becoming large language models, a neural network on steroids, so to speak. They’re much more sophisticated. A prime example is ChatGPT.
Here’s why you need to know about old school ANNs right now.
A retrospective analysis looks back in history. A clinician researcher might look back at the records of a thousand patients to figure out if a particular ANN works. If the data says it works, that’s encouraging. However, in a retrospective analysis, there are many confounding variables that can influence the results.
On the other hand, with a prospective study, a clinician researcher starts looking at today and moves forward in time. It’s a much more reliable form of evidence.
The bottom line: Over the last few years, there have been a lot more prospective studies on ANNs, including the most reliable prospective studies, called randomized controlled trials (more on these below). And – here’s the kicker – this kind of research is convincing many clinicians that ANNs are worth using today.
Paul Cerrato is a senior research analyst and communication specialist at Mayo Clinic Platform, which focuses on finding digital solutions to many of the stubborn problems that face healthcare. He also is a professor at Northeastern University D-Amore-McKim School of Business, where he teaches data mining and machine learning.
Cerrato has been widely published in peer-reviewed medical literature and has written seven books on digital health. He also has been named one of the most influential bloggers by HIMSS. Further, he recently was inducted into Sigma Xi, the Scientific Research Honor Society, which in the past has included Albert Einstein, Linus Pauling and more than 200 Nobel Prize winners.
Healthcare IT News sat down with Cerrato to do a deep dive into artificial neural networks.
Q. What is an artificial neural network, and what does it do?
A. An ANN is a collection of nodes that are interconnected. The word node brings to mind a neuron – thus the term “neural network.” They’re a lot like the neurons or the nerves within the brain.
If you’ve ever seen a picture of a neuron in the brain as a centerpiece with little tentacles going in and out of it – those tentacles are the connections. One neuron has a tentacle coming out of it, connecting to the next neuron and the next neuron, so there’s a network. In many respects, the artificial neural networks are like the neural networks in the brain in that they start with an input, they go through all these different nodes or neurons, and they eventually come out with an output.
The input might be something like thousands and thousands of pictures of skin lesions. Let’s say you have a mole on your skin, and the doctor has to decide if it’s a melanoma or a normal mole. The neural network will collect say 10,000 pictures, and then analyze them through a series of layers, a series of nodes, and the output would be either yes, it’s a melanoma, or no, it’s a normal mole.
Thus, a neural network is a collection of nodes that go from input to output to help decide whatever problem you’re trying to solve.
Here’s an image that will help. [Editor’s Note: See chart above.] On the left, you have tens of thousands of images of skin lesions, and the neural network has to decide what they are. It goes through this analysis process, which is basically a series of mathematical calculations, and then at the end of it comes the conclusion – either it’s a melanoma or it’s not a melanoma.
That’s what a neural network does. It analyzes data, whether it be an image or some other data, and then generates a result.
It starts with a training process. In the training process, let’s say you’ve got these 10,000 images. It can split those images into 5,000 and 5,000. The first 5,000, the network will analyze the data and come up with a solution. This includes what’s called ground truth. In other words, you tell the neural network the correct answer for every one of those 5,000 images.
The network will make a lot of mistakes, but it will go back and fix those mistakes, back and forth, back and forth, adjusting those numerical values until it gets to such a point where it’s almost completely accurate in generating the answer you entered, melanoma or not melanoma.
So that’s the training part. Then comes the testing part.
The second 5,000 images go through a testing process in the neural network. Here you see how accurate the neural network is in answering the question, because in the second 5,000 images, it’s not being told the correct answer. It must figure it out on its own.
Depending on how accurate this process is, you’ll see if the neural network is worth using by clinicians. If the numbers are 80% accurate, then a doctor would be more inclined to use it. But if it’s 30% accurate, they will throw it out and say, “Well, it’s not really going to help me make the diagnosis.”
To recap, during the training process, each of the images is labeled so the network knows the correct answer. Then, during the testing stage, the network doesn’t know the answer. So that’s a blind test that enables them to figure out how accurate the network is going to be.
Q. What is new about ANNs that has convinced you to get the word out on them?
A. One of the criteria we use in healthcare to decide if something is worthwhile is the amount of evidence to support the intervention. Over the last 10 to 15 years, we’ve been writing about neural networks, and a lot of the research has been what’s called retrospective analysis, as opposed to prospective analysis.
The bottom line is retrospective analysis is a much weaker form of evidence. So, there’s less reason for doctors and thought leaders to believe in the results when they’re retrospective, as opposed to prospective. Prospective research today is convincing many doctors that these tools are worth using. [Editor’s Note: See introduction above.]
Artificial neural networks that use machine learning have been used in several programs at Mayo Clinic. One of these ANNs helps physicians doing colonoscopies improve their accuracy. There’s something called an adenoma detection rate, which measures how well the ANN is able to detect precancerous polyps in the colon. This detection rate has gone up significantly since ANNs have been used. It’s saving lives.
Mayo Clinic’s Eagle Study is an example of a randomized controlled study that provides strong support for artificial neural networks. Such randomized controlled studies offer safeguards to help determine whether or not a study’s results are believable.
The Eagle Study used an artificial neural network in combination with an EKG to see if that combination could help detect a weak heart pump, which increases the risk of heart failure. Several clinicians were given this AI, the known network, plus an EKG. And then a separate group of doctors were given the choice to not use the AI.
Those who used the AI got better results. They were more likely to pinpoint those patients who were at risk for a weak heart pump. It’s that strong evidence that has convinced more and more doctors to use ANNs.
A specialized type of ANN that has caught everyone’s attention is the large language model. It typically draws on massive datasets including billions of examples to generate a model (or algorithm) that can help diagnose disease, summarize patient records and answer patients’ emails, if carefully constructed. On the other hand, if they derive their data from unreliable sources, they are prone to errors, and have been known to invent answers, sometimes referred to as hallucinations.
And ANNs and LLMs are not just being used for diagnosis. They’re also being used to summarize electronic health records. A doctor walks into a hospital room and must figure out what to do next for Mrs. Jones. The doctor opens up her electronic health record, and there’s 100 pages of information. There’s no way you’re going to read 100 pages of information. What you want is a concise summary of the most important points in those 100 pages.
So, electronic health records are now being plugged into these artificial neural networks, including the large language models, and then generating a result like the 10 things most important for you to know. If the ANN has been carefully constructed and validated, it can help the doctor make a decision as to what to do next.
So, they’re being used for diagnosis. They’re being used to summarize large documents. They’re even being used to help answer patients’ emails. Doctors now are responsible for answering patients’ emails, and they get bombarded with all these questions, and some of them can be answered through AI. If they’re relatively simple.
For example, what are the side effects of the drug you prescribed? Do I need to come in for a follow-up appointment? An artificial neural network or a large language model can give an answer to that. Then the doctor can look at the answer, and if it’s correct, send it along. Or they can determine it’s not correct and write an answer themselves.
So, there are safeguards in that. The doctor still has to review these things. But it saves a lot of time because the doctor doesn’t have to type them all out himself.
Q. Are you using artificial neural networks anywhere at the Mayo Clinic today?
A. First of all, we’re not at the stage yet where we have a large language model that’s available to clinicians and to the public. We’re working on that. That’s a work in progress. What we do have are artificial neural networks being used, for example, for the diagnosis of colon cancer, as a perfect example.
Take what an endoscopist or a gastroenterologist sees when he or she sticks a scope up into your colon to look for polyps. Usually, if the polyp is precancerous, it needs to be removed. Otherwise, eventually it becomes cancer.
The problem is when you’re looking through a scope, you’re human with human eyesight, and you can’t always detect some of the subtle things that can be detected by a computer. Because a computer’s vision is much more sophisticated. A computer can look at millions and millions of pixels and analyze all those pixels to find something unusual the human eye can’t see.
What you’re seeing is two examples of very small, easily missed adenomas or precancerous polyps. And the computer is highlighting where they are in the colon and basically saying, “You see those little things there? You might think about removing those because they may be precancerous.”
You may not have noticed that with the human eye. So, this is the type of program that’s being used by physicians at Mayo Clinic to improve the diagnosis of colon cancer. And the numbers are really good. The research shows the adenoma detection rate has gone up since ANNs have been used. It’s saving lives.
Editor’s Note: Cerrato points readers to a five-minute PBS clip that simply explains how neural networks work. In this video, it shows how a neural network can figure out what a dog looks like, and what it doesn’t look like.
Follow Bill’s HIT coverage on LinkedIn: Bill Siwicki
Email him: [email protected]
Healthcare IT News is a HIMSS Media publication.
WATCH NOW: Chief AI Officers require a deep understanding of the technologies and clinical ops