Editor's Choice: 10 Trending Stories from November

The rapid rise of AI could potentially change healthcare forever, leading to faster diagnoses and allowing providers to spend more time communicating directly with patients. According to a new report from the Brookings Institution, however, there are also risks associated with AI in healthcare that must be addressed.

These are six potential risks of AI that were identified in the nonprofit organization’s report:

1. Injuries and error: “The most obvious risk is that AI systems will sometimes be wrong, and that patient injury or other healthcare problems may result,” author W. Nicholson Price II, University of Michigan Law School, wrote. “If an AI system recommends the wrong drug for a patient, fails to notice a tumor on a radiological scan, or allocates a hospital bed to one patient over another because it predicted wrongly which patient would benefit more, the patient could be injured.”

Errors related AI systems would be especially troubling because they can impact so many patients at once. In addition, patients and the patients’ family and friends are likely to not react well if they find out “a computer” is the reason a significant mistake was made. And in this modern era of online patient reviews, it would not take long for the word to get out that a providers’ AI capabilities could not be trusted.

2. Data availability: The logistics related to the patient data needed to develop a legitimate AI algorithm can be daunting. Even just gathering all of the necessary data for a single patient can present various challenges. As Price II explained, patients “typically see different providers and switch insurance companies, leading to data split in multiple systems and multiple formats.”

3. Privacy concerns: When you’re collecting patient data, the privacy of those patients should certainly be a big concern. Researchers may work to ensure that patient data remains private, but there are always malicious hackers waiting in the wings to exploit mistakes. Even a massive company such as Google can experience problems related to patient data and privacy, showing that it’s something everyone involved in AI must take seriously.

“AI could implicate privacy in another way: AI can predict private information about patients even though the algorithm never received that information,” Price II added.

4. Bias and inequality: If the data used to train an AI system contains even the faintest hint of bias, according to the report, that bias will be present in the actual AI.

“For instance, if the data available for AI are principally gathered in academic medical centers, the resulting AI systems will know less about—and therefore will treat less effectively—patients from populations that do not typically frequent academic medical centers,” Price II wrote. “Similarly, if speech-recognition AI systems are used to transcribe encounter notes, such AI may perform worse when the provider is of a race or gender underrepresented in training data.”

5. Professional realignment: One long-term risk of implementing AI technology is that it could lead to “shifts in the medical profession.”

“Some medical specialties, such as radiology, are likely to shift substantially as much of their work becomes automatable,” Price II wrote. “Some scholars are concerned that the widespread use of AI will result in decreased human knowledge and capacity over time, such that providers lose the ability to catch and correct AI errors and further to develop medical knowledge.”

(More AI in Healthcare coverage of this specific risk can be read here, here and here.) 

6. The nirvana fallacy: The nirvana fallacy, Price II explained, occurs when a new option is compared to an ideal scenario instead of what came before it. Patient care may not be 100% perfect after the implementation of AI, in other words, but that doesn’t mean things should remain the same as they’ve always been.

Could this phenomenon occur and lead to inaction in the American healthcare system?

RSNA 2019, the world’s largest radiology conference, kicks off at Chicago’s McCormick Place on Sunday, Dec. 1, and promises to include more AI content than ever before. There will be an expanded AI Showcase this year, giving attendees access to more than 100 vendors in one location.

While it’s impossible to experience every single part of RSNA 2019 related to AI—unless you’re a fully functional robot yourself, I suppose—there are a few educational sessions I do highly recommend you attend. They are:

1. “Artificial Intelligence and Precision Education: How AI Can Revolutionize Training in Radiology” | Monday, Dec. 2 | 8:30 – 10 a.m. | Room: E450A

Attendees should make sure they wake up bright and early Monday morning and make their way to this session, which includes three separate presentations on AI and radiology. The first presentation, from moderator Falgun H. Chokshi, MD, examines the “basics” of machine learning and AI, including its limitations and how it might impact medical imaging in the years ahead.

Soonmee Cha, MD, a professor in residence of radiology and neurological surgery at the University of California, San Francisco, will be delivering the second presentation, “State of Radiology Education: Opportunities for Change.” Techniques for training the next generation of radiologists about AI will—and ways to “educate the educators”—will be discussed at length.

The session is scheduled to end with a panel, giving attendees the opportunity to hear tips and advice about AI from five participants who come from a variety of backgrounds.  

2. “Learning AI from the Experts: Becoming an AI Leader in Global Radiology (Without Needing a Computer Science Degree)” | Tuesday, Dec. 3 | 4:30-6 p.m. | Room: S406B

This 90-minute session includes five different presentations aimed at helping radiologists gain a better understanding of AI. Topics include differentiating AI from global radiology and how radiologists can “get a seat at the table” when it comes to developing implementing AI policies.

The always-fascinating Eliot L. Siegel, MD, is sure to be a highlight with his presentation, “Understanding AI: How Does that Black Box (AI) Work?”

3. “Deep Learning in Radiology: How Do We Do It?” | Wednesday, Dec. 4 | 8:30-10 a.m. | Room: S406B

This session puts the spotlight on deep learning, one of the single most important topics for radiologists to study when it comes to AI. Expect a crowded room for this one—get there early Wednesday morning if you want a good seat!

Luciano M. Prevedello, MD, MPH, will be moderating the session and discussing the use of deep learning at Ohio State University. There will also be detailed descriptions of how deep learning has been developed and implemented at both Stanford University and Mayo Clinic.

This, to me, is perhaps the most exciting part of RSNA 2019 compared to past shows in Chicago—the fact that providers now have real experiences to share. At first, a lot of discussions surrounding AI were about its potential and not about what it was actually like to use these groundbreaking technologies. In 2019, however, things have started to finally shift as more and more radiologists throughout the world are getting their hands dirty and experiencing it for themselves.

Need more details? Click here for RSNA 2019's full schedule. 

It is traditionally believed that men and women experience angina—the pain associated with coronary artery disease—in different ways, with men feeling it in their chest and women feeling it in their arms and back. A new clinical trial out of the Massachusetts Institute of Technology (MIT) in Cambridge, however, suggests there are no differences in how angina impacts the two sexes.

Karthik Dinakar, a researcher from the MIT Media Lab, and colleagues presented the results of their trial at the European Society of Cardiology’s annual meeting back in September. The trial leveraged machine learning technology and included more than 600 patients from the United States and Canada who were referred for their first coronary angiogram. Recorded conversations between those patients and their physicians were then analyzed, with the team observing that chest pain was mentioned in 90% of all conversations.

“This work, showing no real differences between women and men in chest pain, goes against the dogma and will shake up the field of cardiology,” co-author Deepak L. Bhatt, executive director of Interventional Cardiovascular Programs at Brigham and Women’s Hospital in Boston and professor of medicine at Harvard Medical School, said in a news release. “It is also exciting to see an application of machine learning in healthcare that actually worked and isn’t just hype.”

“This sophisticated machine learning study suggests, alongside several other recent more conventional studies, that there may be fewer if any differences in symptomatic presentation of heart attacks in women compared to men,” Philippe Gabriel Steg, a professor of cardiology at Université Paris- Diderot and director of the Coronary Care Unit of Hôpital Bichat in Paris, France, said in the same news release. Steg did not participate in the research. “This has important consequences in the organization of care for patients with suspected heart attacks, in whom diagnostic strategies probably need to be similar in women and men.”

Dinakar noted that symptoms aren’t always fully understood following clinical trials because researchers treat “symptoms as check boxes.”

“The result is to isolate one symptom from another, and you don’t capture the entire patient symptomatology presentation—you begin to treat each symptom as if it’s the same across all patients,” Dinakar said in the news release. “Further, when analyzing symptoms as check boxes, you rarely see the complete picture of the constellation of symptoms that patients actually report. Often this important fact is compensated for poorly in traditional statistical analysis.”

Ultromics, a U.K.-based healthcare technology company, has gained FDA clearance for its new AI-powered image analysis solution.

The company’s EchoGo Core system uses AI technologies to automate the interpretation of ultrasound-based heart scans. The vendor-neutral solution calculates a patient’s ventricular ejection fraction, left ventricular volumes and automated cardiac strain.

“Strain has shown to be very valuable in cardiovascular diagnostics and has been demonstrated in published studies to be linked with earlier detection of disease and improved patient outcomes,” Ross Upton, founder and CEO of Ultromics, said in a prepared statement. “Ultromics will be the first to use artificial intelligence for automated strain analysis, which is applicable to 60 million scans per year. Crucially, strain is also becoming reimbursable from January 2020 in the U.S.”

The first clinical trial to investigate EchoGo’s effectiveness began back in 2011. The company said it is now ready for this “next stage of growth” to begin.

“This is an incredibly exciting step towards the future of healthcare, EchoGo will help clinicians make more accurate and informed decisions to improve patient care delivery,” Upton said. “It's truly a watershed moment for Ultromics.”

Radiologists are in a position to demonstrate their value and lead the implementation of AI in healthcare—but keeping up with these evolving technologies is easier said than done.

A new study published in Academic Radiology detailed one way radiologists have been paying close attention to AI in the last two years: an AI journal club built around thoughtful discussions and interactive webinars.

“AI can potentially assist radiologists not only with image interpretation but also with several noninterpretive tasks,” wrote Patricia Balthazar, MD, department of radiology and imaging sciences at the Emory University School of Medicine in Atlanta, and colleagues. “Regardless of how AI will change the future of our specialty, practicing radiologists and radiology trainees alike must prepare and take the lead in this important practice revolution. After all, if we are not sitting at the table where decisions are made, they will be made on our behalf.”

The American College of Radiology Resident and Fellow Section helped launch the AI journal club, forming an advisory group to help brainstorm ideas for the webinars. The first event, “Radiologists as knowledge experts in a world of artificial intelligence,” occurred in December 2017 and reached more than 60 live participants. The club has produced thirteen webinars in total thus far, with titles ranging from “What does deep learning see?” to “A roadmap for foundational research on artificial intelligence in medical imaging.” Sessions are recorded and later uploaded to YouTube.

These educational events are open to anyone—39% of attendees so far have not been radiologists—and the authors said it is vital for the club to attract interest from individuals with a variety of backgrounds.

“Radiologists, computer scientists, data scientists and others with an interest in the topic are all counted among the attendees,” they wrote. “This diversity is important to disseminate skills and take advantage of the AI revolution to improve patient care.”

The sessions help prepare radiologists for the future, covering certain areas that get don’t get much attention during traditional training. Balthazar and colleagues did note that the AI journal club “does not represent a comprehensive AI curriculum,” but it makes a perfect “complementary tool” for engaging radiologists and other specialists.

“Going forward, the AI Journal Club will continue to hold monthly webinars to discuss the most recent developments, papers, and news related to AI/machine learning,” the authors concluded. “Its format will evolve to attend the demands of the audience and adapt as new resources become available.”

Sheba Medical Center (SMC) in Israel has announced that it is working with XRHealth, a healthcare technology company specializing in virtual reality (VR) and augmented reality, to become the first fully VR-based hospital in the world.

This mission, according to a prepared statement, is a significant piece of the facility’s plan to be “a center of innovation that embodies a startup culture and that encourages the hospital’s complete transformation to digital health.”

“With XRHealth, Sheba Medical Center believes we will be able to provide improved training for our facility, along with better and more personalized care for our patients,” Amitai Ziv, director of  SMC’s Rehabilitation Hospital, said in the statement.

“We've identified medical virtualization as one of the technologies that will transform healthcare,” Eyal Zimlichman, SMC’s chief medical officer and chief innovation officer, said in the same statement. “Within this realm, we aim to be a leader in developing new health services based on VR.”

As a part of this new collaboration, SMC now has full access to XRHealth’s VR platform, including its apps focused on cognitive assessment, training, motor function, pain management and more.

Eran Orr, CEO of XRHealth, called SMC’s work to embrace VR “monumental in both the medical and technology industries” in the prepared statement.

AI can provide significant value to radiologists by sending urgent imaging studies to the top of their worklists, according to a new analysis published in Academic Radiology.

It’s one of many ways AI can make a difference, explained author Saurabh Jha, MD, department of radiology at the University of Pennsylvania in Philadelphia. And radiologists don’t even have to worry about the technology making a massive change to their daily routine.

“The algorithm does not do radiologists’ work,” Jha wrote. “The radiologists’ responsibilities and visual and cognitive workload are unchanged. By prioritizing, AI addresses a specific uncertainty radiologists face which is that the importance of studies is generally unknown to radiologists before they see the images.”

Exploring the true value of triage

So just how much value can these triage algorithms provide? It depends on the specific situation. In parts of the world where specialists are in short supply, for example, the algorithm practically stands in as the attending radiologist. The value being provided in those situations is colossal.

“It can take several hours’ travel on undeveloped roads to reach the nearest trauma surgeons,” Jha wrote. “A CT scan may be available locally but no radiologists may be present to interpret the images. If AI can detect life-threatening abnormalities and only positive cases are transferred, regional hospitals, which are scarce resources as well, would not be inundated with patients who do not need treatment.”

The value is more “incremental” in countries such as the United States, where radiologists are readily available to interpret studies as needed, but AI is still making a clear impact in those situations.  

In his analysis, Jha put a spotlight on the push to use AI for triaging head CT scans. Time is brain, as the old saying goes, and every second counts when providers are dealing with a neurological emergency. AI, however, can help ensure that radiologists see the most important examinations right away. By reducing the turnaround times for such important cases, AI is providing considerable value to all parties involved.

The analysis also emphasized that triage algorithms aren’t necessarily always just looking for “the most clinically severe condition.” What the AI can actually reveal is which conditions is “the most reversible.”

“The more time-sensitive the reversibility, the greater the value of triage,” Jha wrote. “Widespread intraventricular hemorrhage, which has a dismal prognosis a priori, is not as time-sensitive as ischemic stroke with subtle gray-white matter effacement in the basal ganglia, which can be reversed by prompt thrombolysis.”

The danger of false-positive, false-negative findings

Of course, as is the case with any other algorithm, triage algorithms aren’t perfect. There can be false positives (FPs) and false negatives (FNs), Jha wrote, putting less urgent studies ahead of studies that actually do need to be read as soon as possible.

One side effect of this is that it can actually cause the radiologist to second-guess their own interpretations.

“Radiologists would not know whether a study is positive or negative until they have looked at the images,” Jha wrote. “But they will know whether AI thinks the study is positive or negative by virtue of where the study is placed in the queue. In AI-triage's absence radiologists may have called a study ‘normal,’ but their presence may encourage radiologists to see abnormalities which do not exist or hedge in their reports.”

Working closely with AI has led researchers from one institution to embrace a familiar piece of technology that may surprise many of their peers: the USB drive. The team wrote about its experience in the Journal of Digital Imaging.

“There are major challenges in Canada when developing novel artificial intelligence algorithms using healthcare data,” wrote William Parker, University of British Columbia in Vancouver, Canada, and colleagues. “Simple tasks such as downloading a file from the internet or sending an email from a hospital computer become as challenging as solving the Rubik’s cube. Firewalls, forgotten passwords, app restrictions and administrator permissions plague our computers to the point of hardly being useful.”

USB drives helped the team address these challenges. One example of the technology’s usefulness was how it helped Parker et al. physically bring a completed algorithm to a hospital so that it could be used. The algorithm, designed to identify pneumonia on chest x-rays, was trained by the researchers on their own computer system. Once it was ready to be tested in a clinical environment, however, they couldn’t proceed.

“The hospital computer firewalls and administrator restrictions limited our ability to install new software,” the authors wrote. “We also could not build a web app (website) because data was not allowed to be used in the cloud. Our center restricted data use to on premise only. Further, it was not convenient for our radiologists to visit our lab as it was a 20-minute drive away from where they practice.”

USB drives “saved the day” for Parker and colleagues. The authors encrypted their entire neural network onto a single USB, making it easy for radiologists to test the algorithm’s accuracy and then provide feedback so that it could be improved. With the .exe file never being installed into an actual hospital computer, “there was no risk of affecting the hospital systems or slowing the system computer efforts.”

This feedback paid off immediately, leading to significant “accuracy improvements” that could one day lead to critical improvements in patient care.

The authors did acknowledge that their process was far from perfect. The algorithm’s efficiency is hurt by not being directly installed onto a hospital’s computer, for example, and it can’t be used for “live interpretation” when it’s left on a basic USB drive. Ultimately, though, the team found that this older piece of technology was able to make a world of difference.

“If these institutional issues resonate with you, a USB stick might be the answer to all of your problems too,” they concluded.

Deep learning techniques have shown potential to change cardiac MRI forever, according to a new analysis published in the American Journal of Roentgenology. However, the authors wrote, it is also important to remember deep learning’s current limitations.

“Quantitative analysis for cardiac MRI has been a much loved topic in medical image analysis not only because of its clinical utility, but also because of its technical challenges,” wrote Qian Tao, Leiden University Medical Center in the Netherlands, and colleagues. “The analysis methods need to tackle the vast variability in cardiac MRI data: the differences in abnormalities, morphology, size and orientation of the heart and also differences in contrast, luminance, artifacts, FOV and signal-to-noise ratio of the image data. Until the recent emergence of deep learning techniques, no classic image analysis method has shown sufficient promise to deal with such a combination of complexity and variability in clinical data.”

Cardiac MRI is used for a variety of purposes, including the evaluation of cardiac structure and function and myocardial scar assessment. The modality “delivers a rich spectrum of information,” the authors explained, and “greatly enhances our understanding of cardiac abnormalities.”

In fact, researchers have found that deep learning can assist providers interpreting cardiac MRI examinations with structure quantification, function quantification, strain and motion quantification, tissue quantification and more. And one of the primary ways deep learning can make an impact on cardiac MRI examinations is by speeding up patient care; even experienced radiologists take more than 15 minutes to interpret a single study, but deep learning can reduce that time to just a few minutes.

“Because the fatigue of manual analysis is taken away, radiologists can focus on more patient-oriented issues such as history and diagnosis,” the authors wrote. “By automating cardiac MRI reading, deep learning can also allow cardiac MRI to be offered at more centers with radiologists with less experience or centers with a high volume of patients and not enough radiologists.”

Deep learning has also shown potential to assist researchers managing clinical trials that require the analysis of thousands of cardiac MRI examinations. The improved precision deep learning brings to the table means that fewer study participants—and fewer laboratory employees—would be necessary, speeding up the entire process and helping make these complex clinical trials more affordable and easier to manage.

Of course, Tao and colleagues noted, deep learning is also associated with certain limitations. A biased algorithm will ruin any attempted research, for example, even if the team behind the algorithm is unaware of any issues. Also, there are limited datasets available at this time that focus specifically on cardiac MRI—and without the right dataset, AI researchers can’t accomplish much of anything.

Overall, however, the authors concluded that deep learning “has shown excellent performance on multiple cardiac MRI sequences and shows great promise for clinical use.”

“Deep learning algorithms can provide useful information to the radiologists and will enhance the value of cardiac MRI in clinical practice and scientific research,” they wrote. “Meanwhile, research effort should be devoted to further improve its generalizability, interpretability and controllability.”

Healthcare technology is constantly changing, something radiologists know all too well. And while some within the specialty have expressed fear or concern over the continued rise of AI, a new commentary in Clinical Radiology noted that it’s all par for the course—and radiologists must rise to the occasion yet again.

“The key is to ensure that AI is used to address important unmet clinical needs, whether in relation to triage and diagnostic tools, improved workflow or sustainability of healthcare,” wrote Andrea Rockall, clinical chair of radiology at Imperial College London. “Applications of AI must primarily serve patient needs, not just what's interesting or feasible.”

A key component of the use of AI in healthcare, Rockall explained, is the presence of “high-quality data that allow an important question to be answered or a successful tool to be developed.” Radiologists must help make sure the datasets used by researchers are effective, working with “patient representatives, computing colleagues and machine learning scientists” every step of the way.

Radiologists can also help pave the way forward for this new technology by developing strategies that connect this high-quality data with the right researchers.

“We must avoid narrow, biased or discriminatory restrictions,” Rockall wrote. “We must ensure that good ideas (or use cases) can be tested on the best available data-sets. We can look to our colleagues managing tissue banks for strategies that ensure strong data access policies.”

AI research must also take place in a clinical environment as opposed to being limited to research labs and single datasets. Other industries may push out AI algorithms and solutions before they have been tested again and again in an array of environments, but such measures can’t occur when patient care is involved. It’s yet another example of the responsibilities facing radiologists as these technologies continue to grow and change.  

“Our broad radiology community has a unique opportunity to engage with this exciting scientific evolution, working with interdisciplinary ‘convergence science’ teams,” Rockall concluded. “Ultimately, this teamwork has the potential to improve patient care, within a sustainable framework. We should not fear the changes involved, but we must engage with them to ensure patient benefit, and a continuing meaningful role for radiologists.”