How artificial intelligence could transform IR
By Melanie Padgett Powers Fall 2018
Your colleague is treating a patient with a liver malignancy. A computer has evaluated the severity of the disease based on the tumor’s characteristics, as well as the patient’s co-morbidities and lifestyle, recommending a treatment approach that is specific to that patient. It is a personalized approach, not population-based like the Milan criteria.
These examples of artificial intelligence (AI) are not yet available in the clinical setting, but they will be in the near future, experts say. AI is the ability for computers to do things that normally require human intelligence, including learning and making decisions.
“I think that many folks believe that there is certainly a point in the future where it is highly likely there will be computers and machines that have incorporated enough knowledge and logic to be able to self-reason, but we’re simply not there yet,” says James Stone, MD, PhD, vice chair of research and associate professor of interventional radiology at the University of Virginia.
AI has the potential to transform health care, providing more personalized treatment, immense data mining and analysis, and better patient outcomes. What might an AI-supported IR of the future look like? Here are three possibilities.
1. Intelligent medical records
Intelligent medical records could read through a patient’s entire medical record for a physician, understanding and pinpointing parts of the history relevant to the current discussion. The program could use knowledge from data mining and research studies to provide an intelligent synopsis, complete with treatment recommendations. This type of advanced computer analysis might connect a few dots that an IR could not realistically do in a short period of time. It would also help ease the paperwork burden.
“Think about all the time that physicians spend doing that rote, routine cognitive task of looking at something and classifying it. Imagine if you had a system that could help you do that in a much more efficient, effective way,” says Justin Ko, MD, MBA, director and chief of medical dermatology at Stanford Health Care, who presented on AI at the SIR 2018 Annual Scientific Meeting. He’s part of the Center for Artificial Intelligence in Medicine & Imaging at Stanford University, which is developing and evaluating AI systems that benefit patients.
Even if that paperwork takes only five minutes per patient now (perhaps a conservative estimate), imagine getting five minutes back for every patient. “Think about what you would do with all that extra time that you saved. If I had five minutes back for every patient I see in clinic, I would spend that time with my patients practicing the art of medicine and building the therapeutic relationship,” Dr. Ko says.
2. Decision-support tools
Decision-support tools like the ones described earlier in the article could help IRs understand “within the context of an individual patient the true risk–benefit ratio of providing or not providing a procedure,” Dr. Stone says.
“We will be able to take all the information that defines the individuality of that patient and have the option to seek assistance from advanced algorithms to better understand the morbidity and mortality that’s specific to that patient,” he says.
A physician practice could also use an automated monitoring tool to determine individual physicians’ yield rates, analyzing why one IR has higher yield rates than everyone else, says Lawrence “Rusty” Hofmann, MD, FSIR, professor and chief of interventional radiology at Stanford University School of Medicine. He compares it to a credit card fraud alert that goes off when you use your card in a different area of the country or buy something that seems unusual.
“We need a similar type of alert when something goes wrong with what we’re doing because, in IR in particular, we’re always innovating,” Dr. Hofmann says. “Is our new innovation actually better than what we did historically? Whether this is a device or a drug or a new technique we’re using, we need to be able to monitor that to understand the impact.”
Dr. Hofmann’s Stanford research is aiming to create an algorithm to detect deep vein thrombosis (DVT) on CT scans at the time of the scan. He is also co-founder of Grand Rounds, which collects data and uses machine learning to help patients find the right physician using factors such as quality, insurance and distance from the patient (see bit.ly/irqgrand).
3. Dose mitigation
AI could potentially provide more information on the risk of continued exposure to ionizing radiation through the use of fluoroscopy. “There are certainly opportunities for bringing advanced computational power into understanding how best to accomplish dose reduction both to the patient and to the practitioner,” Dr. Stone says. AI could help provide real-time feedback on adjusting overall projections for the patient or what the overall exposure to the operator might be and how to mitigate that dose immediately.
Concerns over AI
With the introduction of more machines into medicine, some have questioned whether doctors could one day be replaced.
While AI might take over some tasks, that doesn’t mean IRs won’t be needed. “My fundamental belief is being a doctor is about the doctor–patient relationship,” Dr. Hofmann says. “It’s the empathetic ear and the shoulder to cry on when you’re trying to help a patient get better. No computer will ever replace that. Ever. But what we need to do is use tools that augment and supplement our skills as physicians. I believe it’s going to make us more efficient and hopefully be able to spend more time listening to and helping our patients.”
Dr. Stone agrees that AI should improve productivity, opening up physicians’ schedules to allow them to focus on their more specialized skills. “I think that it’s going to allow for physicians to be more efficient at their day-to-day work and to be able to accomplish a greater quantity of work in a given day,” he says.
Relieving physicians of dictation and paperwork and letting them spend more time with patients would also help ease the burnout many physicians experience, Dr. Ko says.
A second concern about using AI in health care is that some may want to rely on it too quickly or too much, these IRs say. “The biggest problem is that sometimes the hype gets ahead of the reality, and I think in medicine, that takes on particular importance,” Dr. Ko says. If you move too fast, you break things. “While that might be the mantra in Silicon Valley, when you break things in medicine, bad things can happen,” he says. “That’s why it’s important for us to be very thoughtful about how we approach and implement these advances and technologies.”
Health care also has to be careful about not accidentally creating biases. Research has shown that computer facial recognition programs show a bias for white faces based on the data used, not because of ill intent. An AI program built on a dataset of predominantly Caucasian patients might hold an inherent bias. This means it might not be as robust for patients of different races and ethnicities if that potential bias hasn't been actively studied and addressed, Dr. Ko says.
It’s hard to predict when AI will be available in a clinical setting, but experts say we’re looking at the near future, not decades away. Dr. Ko sees classification as the first step in AI in health care—tools that will be able to identify and categorize nodules and tumors.
“I think that there’s been a lot of excitement around AI, and the reality is setting in that there is a lot of work to get to true AI-based clinical workflows from where we are right now,” Dr. Stone says. “We can do things like teach an algorithm how to recognize a pulmonary nodule, but that’s a long way from teaching an algorithm how to read a thoracic CT scan.”
But out of all the medical specialties, IR is primed to take advantage of AI. “It’s simultaneously a very technical and also a very creative specialty,” Dr. Stone says. “And it’s one that certainly lends itself to high-order reasoning and high-order thinking. So I think our space will really see a lot of assistance over the years with some of these advanced approaches.”
“It’s our moral responsibility to be collecting data on our patients in a method that allows us to analyze it and understand our impact on them,” says Dr. Hofmann. “We owe that to our patients, and we now have the ability to do so.”