Diving Into the Business of Big Data in Radiology

A patient goes in for a CT angiography exam. In addition to any pulmonary embolisms, physicians receive his calcium scores, bone density measurements and cardiac chamber size. The patient’s numbers are then compared against millions of similar patients, and an algorithm is used to determine that he has triple the risk for deep vein thrombosis. Although this perfection of personalized medicine is not yet a true story, big data may soon make it reality.

“Medicine is no longer encapsulated in one thing. We need more information than is encapsulated in a population, and have reached the conclusion that medicine needs to be personalized for appropriate decision making,” says Carl Jaffe, MD, professor of radiology at Boston University School of Medicine. Ultimately, the value of big data is tied to its ability to personalize recommendations for an individual. The holy grail of big data would be achieved by assessing a patient’s images, cross referencing findings, correlating them with genomic or histological data and them comparing the whole picture with similar cases to make predictions about treatment and outcomes.

The 3 V’s of Big Data

The amount of data

The speed that data is acquired

The different sources of disparate data

At its core, big data offers the opportunity to identify areas for improvement in any given field by making huge scores of data accessible, usable and understandable. When applied to healthcare, big data seeks a lofty goal: to revolutionize clinical care so that it is more effective, efficient, precise and personalized. In its simplest terms, big data is millions of terabytes of data. The buzzword, however, encompasses much more.

“When I’m talking about big data, I’m not just talking about the number of bytes and bits, I’m talking really about the incredible complexity associated with the image,” according to Eliot Siegel, MD, of the University of Maryland School of Medicine and VA Maryland Health Care System, who spoke at the 2014 annual meeting of the Society for Imaging Informatics in Medicine (SIIM) in Long Beach, Calif.

Big data is particularly important to radiology, a specialty in the business of big information and a key specialty in individualized decision making. Given the changing healthcare climate, radiology departments have to watch their spending due to heightened competition, declining reimbursement and pressure from payers to eradicate unnecessary imaging. It will become increasingly difficult for radiologists to obtain reimbursement and funding if people can’t see the data, and thereby value, in scans and reports. This is where big data comes into play.

“Today, we store images as DICOM files with some metadata, but there are a large amount of visually extractable data within these images, such as location of organs, measurements, annotations and location of pathology,” says Khan Siddiqui, MD, visiting associate professor of radiology at Johns Hopkins University School of Medicine and president and CTO of higi, llc, a healthcare startup based in Chicago aimed at helping consumers keep track of their health. “We’re only touching upon the tip of machine-enabled analysis of medical images.”

This hidden data, also known as pixel data, was likened to dark matter in space by Siegel at the SIIM meeting. Dark matter cannot be directly observed with current technology even though it’s thought to comprise more of the universe than observable matter. A standard CT pulmonary angiogram, he noted, theoretically contains thousands of typically unreported parameters that providers don’t yet have the capacity to parse out.

Data analysis has already led to some personalization of diagnosis and treatment protocols, explained Katherine P. Andriole, PhD, director of imaging informatics at Brigham & Women’s Hospital in Boston and professor at Harvard Medical School, during the honorary Dwyer Lecture at the SIIM. MIT researchers John Guttag, PhD, and Collin Stultz, PhD, built a computer model to analyze EKG measures of heart attack patients that normally would have been discarded. By using data mining and machine learning, the researchers analyzed massive amounts of data to reveal three electrical abnormalities that identified which patients had a double or triple risk of dying from a second heart attack within one year.

Although the use of clinical analytics in healthcare is just beginning, many radiology providers are investing in business analytics to streamline operations and reveal opportunities that could have been missed otherwise. Analytics programs run from billing systems, RIS, PACS and voice recognition systems. Many vendors also are offering turnkey solutions for business and clinical analytics. Big data searches for correlations rather than causes, and patterns rather than sums. Patterns are valuable to observe; if money is saved for a system, then money can be made in return.

Big data 101

Still don’t quite get big data? The nature of big data can be divided into three parts, explains Siddiqui. These components are known as the three V’s: high volume, high velocity and high variety. Taming data on this scale requires new forms of processing to enable enhanced decision making, insight discovery and process optimization.

To address the three V’s of big data, a multi-disciplinary approach is needed. The implementation of big data can be separated into three segments: big data infrastructure, big data science and big data execution. All three parts require different knowledge, sources and resources. The infrastructure refers to what systems are needed to store and process data, while the science component references the tools and knowledge needed to create insights about the data. Finally, big data execution refers to the way in which insights and knowledge are generated from the data and then adopted into clinical practice.

The larger the dataset, the more powerful and diverse the information, which can be intimidating to decipher and understand. Big data doesn’t help anyone if few people can access and comprehend it, explains Jaffe. At the National Cancer Institute, image archives are being made publicly available for this reason. The raw data and treatment decisions made have to be organized, curated and checked for reproducibility and validity.

“Science is a conversation,” says Jaffe. “The National Cancer Institute has made efforts to implement serious data sharing plans. There needs to be movement along a lot of lines to get to the truth.” Big data represents one internally consistent version of the truth, and it is imperative that the data gathered are reliable. If mistakes are made during this process, then analytic outcomes are negatively affected.

Adding to this idea, Andriole suggests an additional “V” be included in the big data equation: veracity. The information at hand must have integrity and validity for accurate, effective analysis, she says.

According to Jaffe, big data is already happening in radiology and has been for some time. “It’s not as institutionalized as it can be, but radiology is ahead of many fields, as there’s a much greater capacity to look at data in the imaging community,” he says.

Matthew Morgan, MD, MS, assistant professor and chief of imaging informatics at the University of Utah School of Medicine and Breast Imager at Huntsman Cancer Institute in Salt Lake City, echoes Jaffe’s contention. “A lot of data have been accumulating in radiology and medicine but not much has really changed in the last five to 10 years,” he says. “We’re just continually adding to it. I do think that there are elements of big data in medicine, but I don’t think it’s quite the same as the rest of the world sees it in things like social media and online marketing. I don’t think we’ve done anything fundamentally different recently to necessarily merit that term. If you compare the data from medicine to the volume of data that is accumulating from Facebook, Amazon or mobile apps, then I think it pales in comparison.”

Building a big data plan

Early adopters of big data need to hire architects, data scientists or informaticists to implement a strategy. This need for specialized staff creates a big opportunity, says Morgan. “You need people who have a drive to understand the data. It’s often true that the industry races ahead and starts to generate a lot of buzz around the capabilities of different software, but there’s no button to create insight. We’re likely going to see a theme: the need for people who have the ability to turn data into a compelling story that allows decision makers to see something that they hadn’t seen before—that’s what’s really going to change and drive decisions.”

“It’s going to be a very good time for informaticists and data scientists,” Andriole said at SIIM. “We will be king and we will be needed moving forward.”

When building a big data plan, the problems that require solving for better insight must first be identified. The data that are needed for proper insight must then be accessible. The right tools for extraction need to be created or purchased, the data need to be loaded and the resources need to be secured for analysis.

Steering Data-Driven Hiring into Radiology

Informational asymmetries influence the job market, making it difficult to predict the performance of applicants. By using workforce scientists, many technology companies are now utilizing an intensive, data-driven approach to hiring. Scientists correlate a broad range of signals to job performance metrics for existing employees to create a model that successfully predicts how an applicant with certain attributes will perform in any given situation.

Predictive modeling could one day benefit radiology hiring, according to Douglas Green, MD, of the University of Washington in Seattle, in an article published on May 1 in the Journal of the American College of Radiology. The hiring paradigm could not only include signals like where an applicant completed his or her residency, but also how many cases he or she dictated during the residency and how he or she performed on any visual cognition, emotional intelligence or working memory tests given during the interview process. While this model is still in its infancy, radiology hiring processes could one day be driven by data.

Certainly, creating a big data strategy takes investment. “Ten years ago, we established a data mining program at the Massachusetts General Hospital (MGH) and we have invested several million dollars in this program,” says James Thrall, MD, chairman emeritus of MGH’s Radiology department. “I don’t think individual hospitals would have to do that because those data mining tools are becoming more commercially available. They probably cost a few hundreds of thousands of dollars rather than millions.”

Though investment is necessary, the time and money spent in creating a big data strategy can reap great rewards. “Savings using business intelligence can come in the form of helping us look at staffing and volume. We can adjust our staffing level so that we don’t have too many people working at the wrong time,” says Morgan. “You might have an instinct to hire someone to fix a perceived deficiency when in reality that could be an expense avoided by working smarter. In one sense, big data allows you to right size your efforts along with your opportunities.”

Savings were reaped at MGH through the use of big data and data mining, illustrating the benefit of enacting a business analytics strategy. “By adopting computer order entry that’s informed by data mining, we have reduced the number of phone calls to the phone center by 500,000 a year. Switching to voice recognition from transcription services saved us about $1 million a year,” says Thrall. “Some of the savings come from small pieces that are hard to measure.”

One of these hard-to-measure instances is exemplified by another scenario provided by Thrall. At MGH, data mining is utilized when a doctor orders an MRI scan. Their program searches the patient’s medical record to determine certain things such as whether the patient has an electrical device such as a pacemaker or any allergies. This is a huge time savings for the physician, but it’s hard to put a dollar value on that, explains Thrall.
Big data is also helpful in understanding the effect of doctor, patient and payer behavior on outcomes, which is an essential factor in improving healthcare.

At MGH, all radiology reports were compiled into an electronic database, enabling analysis of radiologists’ behavior in making recommendations. After analysis, Thrall and his colleagues found that there was an almost 100 percent increase over a period of 10 years in the number of recommendations that radiologists were making. Because the referring physicians were no longer visiting radiologists as often thanks to the RIS/PACS, recommendations were given in the reports rather than face-to-face conversations.

“We never would have been able to detect that change in behavior without the availability of big data,” says Thrall.

Addressing the hurdles

Despite the benefits that accompany the execution of big data strategies in radiology, challenges remain. Human resources and technical hurdles loom large, as data extraction tools are very sophisticated. At MGH, a PhD mathematician was hired because people with substantial expertise in managing large databases are needed.

“It’s very easy to make mistakes in analyzing, especially if inquiries aren’t asked correctly,” explains Thrall. Most analyses in themselves are fairly simple, but it takes immense skill to find answers quickly and consistently.

Other difficulties arise when the necessary data are not accessible and tools or technology to extract the data of interest are needed, says Siddiqui. The resources and infrastructure required will depend on what kind of data are being captured and what one wants to do with the data.
Big data ideas also need to be reconciled with time and the necessary intellectual resources. The more powerful the data set, the more discipline and focus needed for the big data strategy to be truly successful.

Storage is yet another issue associated with big data. “As we start to generate insights through analysis of the image data with other healthcare data, environmental data or sensor data,” Siddiqui says, “we will stretch our boundaries of existing technologies as they struggle to handle the data load for processing all of this information.”

In addition to challenges related to technology and human resources, regulatory concerns exist as well. “Though there is already an effective big data and analytics strategy in radiology, the ability to share information poses an issue,” says Jaffe. “We don’t have a way of achieving social commitment to share that data because of privacy regulations.”

Regulatory concerns are typically linked to individual patients, as the privacy rights of patients are now specifically protected through HIPAA and institutional review boards.

For the full promise of big data to be realized, regulation and technology need to work in the complete interest of the patient. “There is a potential treasure trove of data that can’t be accessed or shared due to privacy concerns,” says Morgan. “I don’t think we’ve found the right balance between privacy and opportunity to discover new things. Hopefully there can be more dialogue around those lines and regulations can evolve to open up opportunities to understand disease across the population.”

Though privacy regulations will remain the responsibility of providers, regulatory measures can actually be met through the use of advanced analytics strategies. Data mining can offer insight on compliance, as tools are able to identify all cases with emergent findings and analyze whether the radiologists involved took proper actions. To do this without tools would be a tedious and almost impossible task. This endeavor is made easy, however, with big data.

One last practical instance from MGH further showcases how knowledge from big data can move a basic concept to a higher quality of care for patients. Any patient who is over the age of 65 and admitted to the hospital needs a Pneumovax vaccine, which is good for seven years. Originally, two full-time nurses were hired for the sole purpose of examining 100 to 150 patient records per day to determine whether or not they were vaccinated. This is now done with data mining tools, consequently “providing data at the press of a button that show that we are in compliance with policies,” says Thrall.

Despite these examples of progress, Thrall says big data is still in its infancy. “From conversations I’ve had with colleagues from other institutions, I would say that 98 percent or more of hospitals have not really begun to go beyond business analytics into clinical analytics with big data,” he says. “We are truly at the beginning.”

Morgan adds, “It’s something that radiology probably hasn’t focused on previously, but as payment models change and the slice of reimbursement for imaging will be in negotiation, it’s important for radiologists to have a kind of story that will proactively paint a picture of what our contribution is to patient care.” For sure, big data offers one large way of cementing radiology’s place in the healthcare decision-making hierarchy.