Editor's Choice: 10 Trending Stories from December

Machine learning models can be trained to predict chronic diseases such as dementia using electronic medical record (EMR) data, according to a new study published in Artificial Intelligence in Medicine.

Approximately 5.7 million Americans have Alzheimer’s disease, the most common form of dementia, and treating the disease is expected to cost $1 trillion annually in the United States by the year 2050. Predicting which patients may have dementia later in life, the authors noted, could help screen individuals earlier than ever before and even delay the onset of the disease.  

After considering numerous machine learning techniques for their research, including support vector machine (SVM) and artificial neural networks (ANNs), the author chose a random forest (RF) classifier. SVM was not as interpretable as RF, and ANNs had a lower accuracy.

“This choice was motivated by several factors that were derived from the literature and from our own preliminary investigation,” wrote lead author Zina Ben Miled, PhD, School of Engineering and Technology at Indiana University-Purdue University Indianapolis, and colleagues. “Namely, RF is interpretable, computationally efficient and can handle a high dimensional space of noisy, continuous and categorical features.”

The authors explored EMR data related to more than dementia 2,000 cases from 15 different facilities in Indiana. For the study’s controls, the team used data related to more than 11,000 dementia cases from 25 different facilities in Indiana. The race and gender of the cases and controls were “similar,” reducing bias among the two groups.

The team extracted certain features from the “prescription (Rx),” “diagnosis (Dx)” and “medical notes (Nx)” sections of the EMR. Relevant ICD-9 and ICD-10 codes were also identified. Separate models using EMR data from one year and three years prior to the onset of dementia were developed for the Rx dataset, Dx dataset and Nx dataset. Those same two models were also developed for a combined “RDNx” dataset that merged the three datasets into one.

Overall, the models for one year prior to the onset of dementia had a higher accuracy, sensitivity and specificity. Models developed using the Nx dataset had the highest accuracy, sensitivity and specificity.

Also, the combined RDNx dataset had the higher accuracy of them all.

“This is an indication that, despite the fact that Nx models have a higher prediction accuracy, some of the Rx and Dx features (e.g., antihyperlidemics) make a significant contribution to the overall accuracy of the combined model,” the authors wrote.

The team also observed that age was “consistently among the top features of all the models for both cases and controls” while race and gender did not appear as top features for any model. Race and gender, then, are “unlikely, according to these models, to be significant predictors of

AI continues to evolve at a rapid pace, with new algorithms and solutions being developed all the time. What kind of long-term impact could these technologies have on patient care? What will radiology—and healthcare as a whole—look like in 2030 or 2035?

At RSNA 2019 in Chicago, we asked numerous vendors to look into their own personal crystal balls and predict how AI will change radiology in another 10 to 15 years. A selection of vendor responses can be read below:

• Imad B. Nijim, chief information officer, MEDNAX Radiology Solutions:

“AI in radiology is just as impactful and transformational to our industry as when we moved from film to digital imaging. AI will permeate clinical informatics across all healthcare disciplines, with radiology having the most potential for innovation. First, AI will empower radiologists and impact the clinical practice by improving quality, reducing administrative burden, introducing a new level of image quantification and more. Second, AI will empower the administrators through the use of workflow models to ensure billing efficacy, procedure integrity and fraud detection.”

• Karley Yoder, vice president and general manager of AI, GE Healthcare:

“AI will radically transform radiology over the next 10 to 15 years, allowing clinicians to be much more efficient, accurate and available to spend time with the care team and their patients. This transformation will come as the industry learns how to efficiently unlock data and unleash intelligence.

• Unlocking Data:  For the potential of AI to be realized in healthcare, the first step has nothing to do with AI. The first step is breaking down decades of entrenched data silos that exist within a hospital setting. Platforms will be introduced over the next five years that effectively aggregate data from across the patient record—imaging, EMR, pathology, lab, patient-generated, genomics, etc.—to enable a clinician (and their AI tools) to see the entirety of the patient story and context. Right now this task is left up to a time-strapped clinician to ‘hunt and peck’ across many different systems to manually compile this view. All this work will be eliminated.

• Unleashing Intelligence: Once you have all the data in one place, the next step is to run world-leading AI across the entire longitudinal spectrum of the patient history. This ability will require sophisticated platforms to build these complex and nuanced AI solutions in a manner that is both ethical and generalizable to an entire global population. Furthermore, once the AI is developed, it will not be enough to release thousands of different AI solutions that each tell part of the story. Instead, there will be ‘AI for AI’ that sits on top of each AI puzzle piece and pulls the various insights into one complete view for the clinician to review and act upon. Finally, this completed puzzle of AI then must be delivered in an invisible manner into the clinicians' existing workflow, leveraging the tools and workflow steps they already use and trust.”

• ​​​​​Gene Saragnese, CEO, MaxQ AI:

“In radiology specifically, I think AI is moving us from a place where radiologists are interpreting images to one where they are diagnosing patients. That community does have a history of bringing IT solutions together to achieve a goal.”

• Woojin Kim, MD, chief medical information officer, Nuance:

“Taking advice from Amara's law that says, ‘We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run,’ I hope that we would have made significant advances in AI in healthcare in 10-15 years. I honestly do not know what the impact will be 10-15 years out, as no one predicted the impact of deep learning on radiology today 10 years ago. In healthcare IT, things move slower. There are many challenges when it comes to AI in healthcare, but what I can see for the short term is that more and more research will be done on the evaluation of commercial AI models as more sites will pilot these AI models.

More places will also implement and use AI models in the future. People will use advanced analytics to expedite the validation process and use it for ongoing surveillance as more and more realize its importance and become more familiar with topics like concept drift and data drift. The FDA will hopefully have a mechanism to clear continuously learning algorithms safely.

Over the next several years, I think you will begin to see AI models performing more complex tasks, including identifying more than one finding. As there is greater adoption, I hope these models can move to demonstrate actual patient outcome improvements—even at the personalized medicine level—but that data will take time to gather. I hope this excitement around AI can be a catalyst to update the healthcare IT infrastructure and institutions put even greater emphasis on data governance. I also hope this can lead to some real positive changes in interoperability that is so lacking today. On a more global scale, I hope this technology can be used to address physician shortage issues in underserved countries over the next decade.  

We have to be careful not to overhype and overpromise; as long as we do not enter another AI winter, this progress should continue.”

• Elad Walach, co-founder and CEO, Aidoc:

“In the short term—say, the next five years—AI will be an augmenter and force multiplier, helping clinicians target their expertise via triage, acting as a ‘second reader’ or automating some routine tasks like measurement.

In the long term, I believe radiology AI will move into new, powerful areas like predictive analytics. This won't just be about making the radiologist faster and more accurate; it will uncover completely new, unknown value from medical imaging and be able to detect new signs of disease or risk factors.”

• Hyun-Jun Kim, co-founder and chief strategy officer, VUNO:

“I strongly believe that we will see AI making greater contributions to revolutionizing the entire aspects of clinical workflows, shifting its role from throughput-based applications to providing more individualized and personalized healthcare to patients. That is, AI-driven approaches will enable more rapid, consistent, robust and objective quantification of abnormalities in radiology, allowing medical professionals to diagnose diseases more efficiently and offer bespoke medical care through identifying the types and timing of medical exams a patient needs. This will ultimately lead to early detection and timely treatment of diseases reducing overall medical costs.” 

• Marcel Nienhuis, vice president of marketing, VIDA:

“Initially, we were seeing a lot of AI algorithms offering detection of suspicious features and anatomical segmentation. Now we are seeing algorithms that go beyond detection to automatically measure or quantify image features. The value proposition of both detection and auto-quantification is largely built around physician efficiency. AI algorithms will continue to deliver on automating the mundane portions of radiology reads. 

Soon, we see a world in which the first step of a radiologist's workflow will be to review AI findings, confirm/reject findings and then continue on to efficiently review the case. Beyond detection/quantification is decision support and predictive algorithms. We are starting to see these in the market and have several examples in our own portfolio. The value proposition of these algorithms goes beyond physician efficiency to bring improved patient outcomes, which in turn, lowers the cost of care and improves quality of life.”

AI is almost certain to have a monumental impact on radiology, but what, exactly, will that mean for radiologists?

Some within the specialty fear the technology will replace radiologists altogether, kicking them to the curb like a Christmas tree on Dec. 26. According to a new commentary published in the Journal of the American College of Radiology, however, radiologists will see considerable benefits from the rise of AI—especially if they don’t let other specialists make all the big decisions.

“If we sit back and do nothing, there is a chance we could be marginalized by AI,” wrote lead author Bibb Allen, MD, chief medical officer of the American College of Radiology Data Science Institute (ACR DSI), and colleagues. “On the other hand, if we play a leadership role in AI development, the best days for radiologists, our specialty and our patients are yet to come.”

Allen’s co-authors were Keith Dreyer, DO, PhD, chief science officer of the ACR DSI, and Geraldine McGinty, MD, MBA, chair of the ACR Board of Chancellors. The three authors explored numerous ways AI could help radiologists on a day-to-day basis. AI solutions will be able to detect “subtle yet critical findings,” for example, helping find things that the radiologist could have potentially missed. The technology will also be able to knock out quicker, more straightforward tasks that often take up a radiologist’s time, meaning each specialist can go on to see more patients. And as the demand for radiology is expected to go up in the near future, giving radiologists more time to do their most important tasks is especially helpful.

Allen et al. also noted that AI “does not have to be perfect to be helpful.”  

“Not all radiologists perform identically on every case we interpret, and for that matter neither will AI,” the authors wrote. “If AI recognizes abnormalities not identified by all radiologists and at the same time all radiologists also find abnormalities not recognized by AI, then the combination of humans plus AI has great potential to improve care.”

Another key point of the authors’ commentary was that it will still be a long time before AI is ever trusted with the big “should we” decisions associated with “the human side of medical care.” Radiologists make “should we” decisions all the time, and there are no signs that an AI algorithm is close to being ready to make those decisions on its own.

In addition, Allen and colleagues explained, it seems misguided to simply view AI as a tool to get rid of radiologists. Instead, it’s critical to think of how these advanced algorithms can make the biggest possible impact on patient care.

“To be most impactful, AI developers should focus on the tasks that radiologists cannot do well, such as integrate more information from a variety of sources into our radiological reports, or predict a cancer phenotypic responsive to therapy based on information in the image data that are visually perceptible,” the authors wrote. “Also, algorithms that increase department efficiency and ensure patient safety, including triage of cases with critical imaging findings, will also be immediately useful long before the applications that provide autonomous care.”

The three authors concluded on an optimistic note, noting that there will also be a “need for well-trained radiologists.” And AI may lead to some significant changes, but there is no reason to think that those changes will lead to radiologists being replaced.

“Predicting future workforce needs has always been challenging, and predicting our workforce as AI evolves may be no different,” the authors wrote. “But the future seems incredibly bright for our specialty.”

AI could change healthcare forever. But for the technology to reach its full potential, researchers must make sure they go about things the right way, according to a new report from the National Academy of Medicine (NAM).

The report, “Artificial Intelligence in Health Care: The Hope, the Hype, the Promise, the Peril,” offers readers an in-depth examination of the current state of AI in healthcare. The 245-page report, featuring contributions by 27 different authors, can be read in full on the NAM website. Michael E. Matheny, MD, MS, MPH, of Vanderbilt University Medical Center and the Department of Veterans Affairs, and Sonoo Thadaney Israni, MBA, of Stanford University, served as co-chairs for the project.

Matheny and Israni also worked together to write an opinion piece for JAMA covering many of the report’s findings. Danielle Whicher, PhD, MHS, of Mathematica Policy Research—who was not one of the original report’s authors—was a third co-author of the JAMA commentary.

“The promise of AI in healthcare offers substantial opportunities to improve patient and clinical team outcomes, reduce costs, and influence population health,” Matheny et al. wrote in JAMA. “Current data generation greatly exceeds human cognitive capacity to effectively manage information, and AI is likely to have an important and complementary role to human cognition to support delivery of personalized healthcare.”

However, certain challenges remain when it comes to realizing AI in healthcare’s full potential. According to Matheny and his co-authors, these are three crucial points that all AI stakeholders—researchers, data scientists, physicians and vendors alike—must keep in mind for AI development and implementation to be successful:

1. Algorithms must be trained and validated on population-representative data:

More healthcare data is available right now than ever before—and it’s not even close. But the quality of data being used to develop AI algorithms is still often not what it should be.

“The current challenges are grounded in patient and healthcare system preferences, regulations, and political will rather than technical capacity or specifications,” the authors wrote. “It is prudent to engage AI developers, users and patients and their families in discussions about appropriate policy, regulatory and legislative solutions.”

On a related note, the team also highlighted the importance of “scrutinizing the underlying biases” of AI algorithms well before they are deployed in a clinical setting.

2. Remember that the current focus is augmented intelligence, not fully autonomous AI:

Matheny et al. noted that today’s researchers are working to support healthcare providers and not completely replace them. The public—and physicians themselves—should be reminded of this whenever possible to help avoid any sort of public backlash related to uncertainty about the technology.

“Focusing on this reality is essential for developing user trust because there is an understandable low tolerance for machine error, and these tools are being implemented in an environment of inadequate regulation and legislation,” the authors wrote.

3. Put effective training and educational programs into place:

Over time, these technologies will have such a significant affect on healthcare that proper training and educational programs will be a necessity.

“The curricula must be multidisciplinary and engage AI developers, implementers, health care system leadership, frontline clinical teams, ethicists, humanists, patients, and caregivers,” the authors wrote. “Each group brings much-needed perspectives, requirements, and expertise.”

Consumers will also need to be informed about “consent, privacy and healthcare AI savviness,” the team added.

Eko, a San Francisco-based healthcare technology company, announced that its ECG-based algorithm for heart failure has been granted a Breakthrough Device designation by the FDA.

The algorithm evaluates 15 seconds of ECG data taken from Eko’s DUO digital stethoscope, searching for any possible signs of reduced left ventricular ejection fraction. Such findings are seen as a sign of heart failure.

“The Breakthrough Device designation recognizes the vast unmet clinical needs in identifying heart failure early in patients, whether it be due to cost, inaccessibility, or misdiagnosis,” Connor Landgraf, co-founder and CEO of Eko, said in a prepared statement. “We look forward to working with the FDA to bring this algorithm to patients and to give clinicians a new tool to screen for low ejection fraction.”

The algorithm has already been investigated by a team of researchers from Mayo Clinic in Rochester, Minnesota, with the results being published in Nature Medicine in January 2019.

“A low ejection fraction means that the heart pump is weak, which can lead to shortness of breath, swelling, exercise intolerance, or sudden death, so it is important to identify, as many treatments exist,” Paul Friedman, MD, chair of the department of cardiovascular medicine at Mayo Clinic, said in the same prepared statement. “This technology gives physicians a tool to detect heart disease earlier, and before it develops into a more serious illness. In effect, by imbedding the technology in a commonly used clinical tool—the stethoscope—all caregivers carry some of the diagnostic prowess of an expert cardiologist with them.”

A new imaging technique that uses deep learning technology can identify tumors in colorectal tissue samples with 100% accuracy, according to findings published in Theranostics. 

The study’s authors tested their technique, pattern recognition optical coherence tomography (PR-OCT), on 26,000 optical coherence tomography (OCT) images, achieving a sensitivity of 100% and a specificity of 99.7%. OCT is typically used to capture images of a patient’s retina, but researchers have started evaluating its efficiency in other areas as well.

Senior author Quing Zhu, PhD, a professor of biomedical engineering at Washington University in St. Louis, and colleagues hope their technique can serve as an “optical biopsy tool” for physicians in the near future.

“We think this technology, combined with the colonoscopy endoscope, will be very helpful to surgeons in diagnosing colorectal cancer,” Zhu said in a prepared statement. “More research is necessary, but the idea is that when the surgeons use colonoscopy to examine the colon surface, this technology could be zoomed in locally to help make a more accurate diagnosis of deeper precancerous polyps and early-stage cancers versus normal tissue.”

“The unique part of our system is that we could detect a structural pattern within the image,” lead author Yifeng Zeng, a biomedical engineering doctoral student at Washington University, said in the same statement. “Using OCT, we are imaging something that we can find a pattern across all normal tissues. Then we can use this pattern to classify abnormal and cancerous tissue for accurate diagnosis.”

The full Theranostics study can be downloaded at this link.

Researchers have developed an AI algorithm that can identify cancer cells in digital pathology images, sharing their findings in EBioMedicine.

“As there are usually millions of cells in a tissue sample, a pathologist can only analyze so many slides in a day,” Guanghua Xiao, PhD, a professor at UT Southwestern Medical Center in Dallas, said in a prepared statement. “To make a diagnosis, pathologists usually only examine several ‘representative’ regions in detail, rather than the whole slide. However, some important details could be missed by this approach.”

Xiao et al. hope their algorithm, called ConvPath, can help bring about significant changes in that paradigm. Using tumor images from four different datasets, the “cell type classification pipeline” was trained to automatically perform several segmentation and classification tasks on digital pathology images and turn each image into a “spatial map” for the pathologist.

Overall, the team achieved a classification accuracy of more than 90% with both training and testing datasets, noting that their algorithm could “quickly pinpoint the tumor cells” for pathologists.

“It is time-consuming and difficult for pathologists to locate very small tumor regions in tissue images, so this could greatly reduce the time that pathologists need to spend on each image,” Xiao said in the same statement.

ConvPath could also help pathologists and clinicians “predict the patient prognosis” and “tailor the treatment plan of individual patients,” the authors wrote. And its detailed analysis “could potentially provide information for patient response to immunotherapy.”

To assist other researchers looking to explore ConvPath, the authors shared all source scrips for the software here.

Deep learning models can improve the detection of attention deficit hyperactivity disorder (ADHD), according to new findings published in Radiology: Artificial Intelligence.

Diagnosing ADHD can be challenging, the study’s authors explained, and brain MRI scans could potentially help researchers address those challenges. The connectome—a comprehensive map of a patient’s brain—is a key component of  brain MRI’s potential for detecting ADHD. And improved diagnoses could help more patients get the care they need beginning at an earlier age.

The authors aimed to detect ADHD using multiple connectome maps at once, using existing data from more than 900 patients who were treated at one of eight facilities. All patients had no history of psychiatric, neurologic or medical disorders—other than ADHD.

The team’s multichannel deep neural network (mcDNN) was trained to detect ADHD using both brain connectome data and the patients’ personal characteristic data. Overall, one model achieved an area under the ROC curve (AUC) of 0.82. Single-channel deep neural networks, meanwhile, achieved AUCs of 0.67, 0.69 and 0.77.

“Our results emphasize the predictive power of the brain connectome,” senior author Lili He, PhD, Cincinnati Children's Hospital Medical Center, said in a prepared statement. “The constructed brain functional connectome that spans multiple scales provides supplementary information for the depicting of networks across the entire brain.”

The team’s technique isn’t limited just to ADH, He noted.

“This model can be generalized to other neurological deficiencies,” she said in the same statement. “We already use it to predict cognitive deficiency in pre-term infants. We scan them soon after birth to predict neurodevelopmental outcomes at two years of age.”

A majority of U.S. consumers are confident AI will have a positive impact on both patients and healthcare providers, according to a new report from Blumberg Capital.

The report, which included survey responses from 1,000 adult consumers from the United States, found that 81% of consumers think AI will directly improve patient care. Eighty-two percent, meanwhile, think it will lead to more collaborations between physicians and healthcare facilities. In addition, 80% of respondents think AI will “have a positive impact on the healthcare workforce itself.”

“The potential of AI in the healthcare industry to deliver major impact is here and now,” according to the report. “Consumers are increasingly open to and, in fact, demanding that care providers, insurers, hospitals and the entire industry adopt new technologies such as AI that can improve healthcare, reduce costs and minimize their concerns about the safety of their bodies and their data.”

Concerns do still exist

Respondents were less optimistic when it came to the performance of advanced AI algorithms. While more than half of consumers think AI could cause errors, 34% worry about AI leading to the mismanagement of their healthcare. Another 18% said they worry about AI-powered robotics systems performing “botched” surgical procedures.

Twenty-five percent of consumers are concerned AI could lead to data security issues, a number that seemed low to the team at Blumberg Capital.

Deep learning (DL) can predict and enhance MS lesions on unenhanced MRI scans, according to a new study published in Radiology. These findings suggest MS patients may someday be treated without the use of gadolinium-based contrast agents (GBCAs).

“There are concerns about GBCA administration, including nephrogenic systemic fibrosis in patients with renal compromise and long-term gadolinium deposition in various tissues,” wrote lead author Ponnada A. Narayana, PhD, MSc, University of Texas Health Science Center in Houston, and colleagues. “This is particularly a concern in patients with MS, who undergo frequent imaging with GBCA administration for regular clinical follow-up, which may result in higher cumulative gadolinium deposition in tissues.”

The team’s DL model included two networks, a convolutional neural network (CNN) that scanned all 2D slices for possible contrast enhancement and a fully connected network that created a participant-wise prediction.

The authors then performed a prospective analysis of MRI data from more than 1,000 adult patients with relapsing-remitting MS who were recruited to participate in a multicenter phase III clinical trial from 2005 to 2009. Sixty-eight different facilities contributed to the original dataset.

Overall, at least one enhancing lesion was detected in 519 study participants. Averaging a total of five test sets, Narayana et al. wrote that the sensitivity of their model for slice-wise prediction was 78% and the specificity for slice-wise prediction was 73%. The sensitivity and specificity for participant-wise prediction were 72% and 70%, respectively.

“The reduced accuracy for participant-wise prediction (70% compared with slice-wise accuracy of 75%) is not surprising; errors in slice-wise predictions are amplified and may completely alter the prediction of patients without enhancing lesions (resulting in a false-positive result) or patients with a single enhancing lesion (resulting in a false-negative result),” the authors explained.

Also, the average area under the ROC curve (AUC) for slice-wise prediction was 0.82. For participant-wise prediction, the AUC was 0.75. These findings, the team wrote, show that DL could someday make the use of GBCAs for the identification of enhancing lesions unnecessary—but more work still needs to be done.

“It is anticipated that inclusion of other MRI sequences could further improve DL performance,” the authors concluded. “This will be necessary before DL is accepted as a viable alternative to GBCAs for identifying enhancing lesions in MS.”