I had this question in my inbox for a while. “What do you think about artificial intelligence in medicine?” Well, a month ago I would have scoffed at the idea. Then I interviewed Dr Richard Do, prominent radiologist at MSKCC for Radiology Today. He said something that completely shifted my mindset.
“It’s all about data,” he said. “A single pathology slide has far more data than an entire CT. The more data you have the better that machines can work with it.” He’s right. Most medicine is data. What we do is take data we know and extrapolate it to current conditions. That’s what clinical trials are and that’s exactly what we can train artificial intelligence to do.
More of us need to start thinking like him. In many areas today, research has stalled at the genetic level. At present, it’s very difficult to treat at this level even if we make astonishing discoveries. Most of our macro level issues are patient compliance, longer wait times, the dearth of doctors, having untrained people prescribe and administer healthcare. Just yesterday, The Daily Mail reported that pharmacists were going to dispense medication due to the unavailability of staff in the NHS. We know how pharmacies operate and in the end it’s going to end up being about the money.
So Artificial Intelligence seems to be a good option at the moment.
ARTIFICIAL INTELLIGENCE IN RADIOLOGY
Radiology is already considering it’s next generation tools in AI. Whethere that’s patterned learning through texture or radiomics, they’re now trying to study virtual biopsies. How? In the area of cancer research, AI can study each pixel of tumors and then determine the malignancy potential through a series of mathematical calculations and algorithms. There’s already been some success with this in colorectal and breast cancer.
Machine learning and deep learning can help diagnose and dictate canacer treatments. This is going to get better with increasing datasets i.e more CT’s, MRI’s, mammograms and ultrasounds. Although, with neural networks, data shows that with multiple images the networks can sometimes “hallucinate.” This can be dangerous and requires a lot more work before application. I’d love to explain all this but I’m not a radiologist. So if you want to know more about radiomics, you’ll have to check this out.
Similarly, a research scholar from the University of Washington won the Paul Baran Young Scholar Award for 2018. She used smartphones to detect life-threatening symptoms. The idea came from bats who use sonar to detect in the dark. Using the very same principles of acoustics, movement and respiration can also be tracked via smartphones. AI in radiology looks promising. Every year there are plenty of new ideas coming to the fore. And with not a lot capital required, radiologists have warmed to the idea themselves. In fact, smaller tasks like measuring changes of lung nodules or size of multiple sclerosis lesions may be automated in the future.
ARTIFICIAL INTELLIGENCE IN OB/GYN
In OB/GYN big data has already positively impacted to breast and ovarian cancer. However, artificial intelligence for detecting preterm uterine activity successfully and accurately identified spontaneous preterm labor. Doctors used electrohysterography signals to detect preterm births, as they provide a strong basis for objective prediction and diagnosis of preterm birth. They used three different machine learning algorithms to identify women with premature births in the records. The results revealed that the Random Forest Algorithm performed the best of sensitivity 97 and specificity of 85%. Area under the Receiver Operator curve (AUROC) of 94% and mean square error rate of 14%.
Similarly, advanced AI-powered EFM with push-button reporting can notify clinicians about patients when conditions are worsening. Early-warning systems, like them are highly effective as they continuously analyze fetal heart rate, contractions, and labor progression, and maternal vital sign alerts. A study showed that with the Maternal Early Warning Trigger (MEWT) tool reduced maternal mortality. Particularly, in these 4 most common areas of maternal morbidity: sepsis, cardiopulmonary dysfunction, preeclampsia-hypertension, and hemorrhage.
ARTIFICIAL INTELLIGENCE IN PATHOLOGY
There was an article a couple of years ago about Pathology and AI becoming frenemies. In the end, it concluded that AI and Pathology were natural co-operators. No one is talking about replacing pathologists. But in the last few years, it’s been clear that AI can improve the productivity of pathologists. Google’s LYNA algorithm proved it could increase speeds in analyzing stained lymph node slides and sensitivity of micrometastases detection in recent studies. Pathology is rich in datasets.
In another project, Dr. Lundin, associate professor for biomedical informatics and research director at the Institute for Molecular Medicine Finland (FIMM) makes the diagnosis directly from an image for risk prediction. No pathologist was needed. They trained a classifier, as in cancer, with images of tissue and used patient outcome as an endpoint in the AI training process. The algorithm machine simply learns the patterns of images, but outperforms experienced pathologists in predicting patient outcome. How cool is that? 🙃
ARTIFICIAL INTELLIGENCE IN SURGERY
In surgery, I’m not so sure. AI in surgery makes me immediately
think of robotics. The way I see it, it’s too hard to train, too many hours to train and hard to research. I mean come on, where can you practice except on a live patient. Who’s going to be the guinea pig here? For micro level procedures, “robotic hands” are great. A procedure done by a robot to suture small blood vessels, no larger than .03 mm – .08 mm across is a good example. This study at the Maastricht University Medical Center in the Netherlands used robotic hands with great success. There’s talk of using it to teaching future surgeons and showing them which techniques are associated with better outcomes. Overall, it’s a nay for me.
ARTIFICIAL INTELLIGENCE IN PSYCHIATRY
In a landmark paper in JAMA Psychiatry, Adam Chekroud, Ph.D., an assistant professor of psychiatry at Yale and colleagues used machine learning in patients with depression. They analyzed the predictive value of 20 depressive symptoms on choice of antidepressant medication. Psychiatrists used the data of more than 7,000 patients from nine antidepressant clinical trials. They discovered three “symptom clusters” that predicted response to antidepressant medication. Similar studies for schizophrenia and the doctors could predict good versus poor treatment outcome. That still doesn’t leave out the fact that human connection is still important when it comes to psychiatry. And so where does machine learning in this field leave the doctors?
PROBLEMS WITH AI
Data: These algorithms generate A LOT of data. Who’s to say your algorithm is better than mine? Or that the algorithm from Harvard is better than the one from Yale? How do we know these artifacts we see on a CT or a slide or a computer analysis are universal? How do we know it’s not just confined to a particular subset.
Sharing Data: This brings me to the next point. If doctors all over the world hold on to their algorithms like it’s their prized golden goose, how can we find out if the findings are universal. There needs to be some body or organization that governs and protects doctors’ and researchers’ algorithms.
Consensus: Until a governing body does emerge for all the information generated, how can consensus be achieved. Pathologists for example, have different end points and how can we agree on them. Many pathologists are still in favor of using older classifications. Some want to rename or reclassify findings as they emerge and the field changes.
Privacy: By far the biggest issue with machine learning and big data is privacy. We’re already seen issues with the hacking of Electronic medical records. Everyone’s worried about it. And with machine learning, it’s only going to get worse as more data goes into the system.
Accountability: Let’s say we do take care of all the issues above. If something goes wrong as things often do in the case of machines, who’s the fall guy? Who’s blamed for a patient’s death or for the wrong diagnosis? Who’s accountable for the error?
My Take
My take on AI is that it’s already here and it’s going to make life easier for those who embrace and learn it, if it’s going to be installed at your institution. Wearable devices, smart phones, apps and tele-medicine is going to become instrumental as more people move online and healthcare moves to cloud. Personally, I think we’re transitioning, certain fields are moving faster than others like pathology. The sooner you get familiar with it, the better it will be. Just err on the side of caution, no machine and no person is perfect. Check and double check, verify and cross verify, and as is with computers, always back up everything.
This article is in no way a reflection of all that artificial intelligence is capable of doing in medicine. If I could, this article would be ten pages long and you’d be asleep by now. 😴 I’ve given you a glimpse of what can be done and shared a ton of links. Whatever your field is, find out how AI is changing it. See if you can contribute or work with these pioneers. Then go and change the world where you live!
What do you make of AI? How it helping your field of medicine?