Dr. Ignacio H. Medrano discusses the role of artificial intelligence in healthcare. AI-powered algorithms have a lot of potential use in medicine. Machine learning improves by patterns and examples rather than by programmed rules. Image-based diagnosis, genomic analysis, and gleaning information from digital medical records are all used today. An important discussion about what’s already present as well as what’s possible.
Speaker: Dr. Ignacio Medrano, Neurologist, Chief Medical Officer of Savana
Let’s get started right away with Google Translate, as it’s a great example of machine learning, which can translate all the languages in the world, even though it doesn’t understand a single one and has no concept of syntax or grammar. How is this possible?
- Google Translate learns the way that children do, by cases.
- Learning by example is much more valuable and practical than going to a language class in many ways.
- Google Translate is an example of world-changing technology in the last 7-8 years.
- Until 2012-2013, all we had were computers that we could program. We are somehow able to “teach” them, which means learning not by rules but by patterns and a ton of examples.
Traditional programming works by teaching rules. Learning by rules has a big problem. Whenever a new case is presented, the machine fails. When you teach it by showing cases, it keeps improving with each example, and we can feed it millions, billions. Even in the real world, most of the time at the hospital, we don’t apply rules. We use what we call intuition, and it’s very efficient.
Learning by cases isn’t without problems. There is a bit of a black box. Google Translate can provide correct answers most of the time, but they cannot explain the why behind the outcome that they provide. Now, let’s look at what machine learning is and what kind of problems it sets out to solve.
Deterministic – You know that you know (no risk). Laws of physics are an example.
Probabilistic or stochastic – Certainty of Risk – You know that you don’t know. Most phenomena fall under the category of risk, but we know what they are, like the lottery.
Heuristic – Intuition machine or machine learning – There is a third counterintuitive category where we don’t know what we don’t know. There is uncertainty about the risk itself.
In that third category, where the risks aren’t clearly defined, intuition or heuristics is a rational way to proceed. We do this every day in medical practice, and we do it when we learn our native language as children. With a computer, we throw millions of cases, not knowing why, but it’s good enough; it works. We’re now creating machines that mimic our intuition. That is what we call artificial intelligence, machine learning, deep learning, or neural networks.
Set of solved problems → Inference of rules → Anticipation of unseen problems
Machines are already set up to infer problems, but now they can anticipate unseen issues, even when it’s the first time they see it. 2012-2013 is where this technology started to make waves in the hubs like Boston, Silicon Valley, Tel Aviv. Very quickly, it got adopted by larger institutions. Open the headlines, and you will see machine learning outperform humans at various endeavors.
Let’s take a few real-world examples of machine learning.
Amazon – They were not Walmart. They were not the best at retail, but they had so much data that they could build a machine learning system to predict who, how, and when people were going to purchase.
Netflix – They were not the best video makers in the world, but they also had a ton of data. Machine learning can plot out factors that will make a film sell.
Machine learning applies to football, politics, fashion to every industry. Machine learning or mathematical intuition is disrupting the world. Broadly, machine learning does two things: Classification and Prediction. They can classify better than humans (ex. Machine learning reveals how much of a Shakespeare play was written by someone else). Prediction adds a timeline and forecast.
Machine Learning and Healthcare
I wanted to apply this to healthcare. Initially, my colleagues were understandably skeptical, but each year brought positive data points.
- In 2016, Google DeepMind was able to diagnose diabetic retinopathy better than ophthalmologists (JAMA). In 2018, the FDA approved this, so it’s used in clinical practice now.
- What’s less known is that in the second semester of 2018, they said: Look, doctors, you’ve seen many retinas, tell me what the chance that this is a man or a woman is? I’m a neurologist, and I’d say this is impossible. There’s no sexual dimorphism in the retina.
- But they gave it to a machine learning system, and they were able to classify correctly on 97% of the occasions. Clinically that’s not too relevant. There are easier ways to know the gender of someone, but it turns out that they were able to figure out cardiovascular risk by observing the retina. In the history of scientific literature, there isn’t a single mention that in the retina, there is a key to finding a person’s cardiovascular risk! Machine learning could do it, and they published this in Nature. That day, I realized how impactful this is going to be for medicine.
- How many other correlations are going to be found by these machines?
- Horizontally, how many biomarkers, in correlation or association, can these machines bring up, which our mind won’t be able to?
- It’s about seeing correlations where our human minds aren’t able to. Machines aren’t constrained by cognitive bias or history.
When you apply machine learning to healthcare, classification fits with diagnosis. Classification is relevant to clinicians, but this is more relevant to managers and providers of hospitals. As a clinician, however, what I find genuinely impactful, is prediction and risk scoring. It’s true; we’ve been scoring risk for 30 years. In this case, though, we humans selected the variables with which we built the scores. That’s why they didn’t work very well. With these new systems that don’t use logistic regression or SPSS, the machines find correlations even when we consider them not to be relevant. Sometimes, however, they are valuable and relevant.
Predictive, precision, individualized medicine is coming, and machine learning will take us there.
- The FDA has approved 37 cases of AI outperforming humans across different medical activities such as prognosis, diagnosis, and therapeutics.
- In terms of papers, we’re probably getting 1 or 2 impactful studies about machine learning and medicine every day. It’s important to consider that these aren’t only algorithms created through images but also other data. They were able to predict which antibiotic was going to work better. It has uses in chemotherapies, as well as Crohn’s disease.
- Did it help the coronavirus? In terms of epidemiological vigilance, not much. However, it predicted which countries would be most affected. CT scans could be used to see whether it’s from COVID or not. Tracking genomes of the virus were used in Israel to red flag patients who are likely to be severely impacted should they get COVID.
- In dermatology, images of melanoma were used. Classification is helpful, but if we could predict, for example, which patients will be more likely to respond to ipilimumab, this is much bigger. In dermatology, both classification and forecasting are likely to be relevant.
On this slide, if you look at the faces on the slide, these are fake humans. They never existed; a neural network generates them. It’s commonly used to create fake Twitter profiles.
- We apply this to microbiological tests or mammography. It applies to dermatology. Published in 2019 on Lancet, machines matching humans in diagnosis was one of the notable advances. Images will be a focus.
- Machines can see correlations where we don’t see them. Sometimes they will be counterintuitive and uncomfortable. It will be our job to identify whether it’s clinically relevant or not because these machines cannot make that judgment.
- Sensors are also interesting. Contactless cardiac arrest detecting and wearable monitoring analytics are already FDA approved.
- AI protein folding algorithms and genomics are another area of surging interest. Prices are dropping, and genomic sequencing moves quickly, but there are reasons for skepticism.
The experts in the field will tell you that we don’t know phenotypes, genotypes, and variants, at this point. This is true, but machine learning is likely to help us uncover these faster than expected. Polygenic scores provide another data layer. Today they can predict the risk of dementia, cancer, and cardiovascular disease. Dermatology will be very relevant here as well. Microbiomics is another area where the potential for machine learning is there.
On a technical note, is having more data layers better? What if I input genomics, microbiomics, environmental pollution, and so on?
- Theoretically, yes. In the future, this will likely be the case.
- At present, in my experience, when you put too many layers together, they don’t work as well. Don’t get stuck on the bureaucracy of setting up data clusters. Start with one data layer, then fit it, and get results.
- There are examples of using two data sets in China – images and genomes put together-and predict macular degeneration risk.
I left the best for last, and this is what we do here at Savana. What data has lots of predictability and is readily available? Clinical records. Everyone has clinical records, but they are unstructured data. It’s a natural target for machine learning, but there are challenges. First, we have a classic computing problem called the free text problem.
- We write in a very unstructured way. Only 15-20% of the information in electronic health records is structured data. As doctors, medical records are often a free train of thought in unstructured free text.
- What if we build technology that transforms this free text into a database? We can then feed the data to machine learning. We would solve a huge bottleneck. So much information is produced every day at hospitals all around the world.
- I develop clinical natural language processing. We’re translating the text in the EMRs, of course, anonymized, respecting compliance and privacy. With that database, we can fit the predictive algorithms. In 2013, almost no one was doing it, but if you look at 2018, 2019, we see how the big academic science departments in the US and China were describing and phenotyping disease in natural language.
- Our team is doing this in a multicentric way, across different hospitals, EMR systems, and languages. We can do this today in English, French, German, and Spanish. Real-world data at the multi-country level, extracting the information directly from the clinical notes, is something that we can’t do manually due to speed constraints.
- Mathematicians, programmers, doctors are all coming together to work on studies across specialties. I invite you to participate as a dermatologist.
- As you can see, we have partners from allergology, cardiology, endocrinology, gastroenterology, hematology, rheumatology, hepatology, immunology, internal medicine, nephrology, neurology, ophthalmology, oncology pediatric, psychiatry, respiratory, and traumatology, and orthopedics. Still, we haven’t started with dermatology yet.
- This might be because you go directly to the images in dermatology, but images are just one data layer. We can add another layer or even combine them.
Changes in Medicine
Let’s look at some general trends in medicine.
We are moving to something more individual, thanks to mathematics but challenges also remain.
- With predictive models, sometimes, they don’t work. For example, with diabetic retinopathy investigation in the US, initially, it was working, but they weren’t working in other countries.
- The only way to see this through is with clinical trials. AI or not, you can only look at the data, study them and find out more, like anything else we do in medicine. Until you put this into clinical trials, everything I’ve said doesn’t count.
- I haven’t proven that this is going to change clinical healthcare. I’m just saying that it’s extremely likely that it’s going to happen.
- The truth is that the clinical trials aren’t ready yet. They are ongoing right now. I have to acknowledge that all of these randomized trials of AI Deep Neural Networks in Medicine are coming from China, which gives you an idea of where this is moving.
The University of Duke in the US decided which faces should receive EEG (electroencephalography) based on risk for seizures. Instead of using classical statistics or SPSS, they gave it to machine learning. The machine started with a few variables, then neurologists jumped in and said, ok, I’m not going to use this and that because the machine is failing here. So this is an excellent example of combining capabilities. Three years later, this resulted in a 63.6% reduction in the duration of EEG monitoring per patient, saving $1135 per patient.
- Any cEEG Pattern with a Frequency of 2Hz
- Epileptiform Discharges
- Patterns include (LPD, LRDA, BIPD)
- Patterns superimposed with Fast or Sharp Activity
- Prior Seizure
- Brief Rhythmic Discharges (double points)
Risk: 0 (less than 5%), 1 (11.9%), 2 (26.9%) 3 (50%) 4 (73.1%) 5 (88.1%) 6+ (95.3%)
This is the first case where machine learning showcased a concrete result in the clinical realm.
Lastly, we need to discuss ethics. When we talk about technology, we also need to talk about ethics and regulation.
JAMA published a paper from China for publication, a bit before the pandemic. It looked at photographs of citizens, and they were able to know who had atrial fibrillation, which is remarkable. I caution that today it’s photographs and arrhythmia; tomorrow, it’s going to be photographs and psychosis or intelligence, perhaps even the intelligence of your children. I think this is something we’ll have to consider.
When you get confronted with machine learning for the first time, you understand its power. Typically as clinicians, we want to have all the answers the first day, and sometimes we get paralyzed by analysis. My humble advice would be, “Let’s isolate the clinical question as if we had to do this manually, and then we call the AI people, and they can build it for us.”
As Picasso said, computers are useless. They can only give us answers. I don’t know if Picasso would have guessed that in 2019, a machine at MIT would be able to ingest all his paintings and generate a not-previously existing but Picasso-like new painting like the one you have here.