Getting beyond the hype around artificial intelligence: 9 key takeaways from industry experts Insight Article Business Intelligence Business Operations Technology Sign in to save MGMA Staff Members Worldwide spending on artificial intelligence (A.I.) was projected to reach $35.8 billion in 2019, a 44% increase from 2018.1 That number is just one of the ways to quantify how A.I. and machine learning have captured the attention of numerous healthcare industry leaders for their potential for innovation. As Mike Cuesta, vice president of marketing, CareCloud, noted at MGMA19 | The Annual Conference in New Orleans, the headlines about A.I. disrupting and transforming aspects of our work and personal lives are rampant. Cuesta joined his colleague Josh Siegel, chief technology officer, CareCloud, and Michael Muelly, MD, product manager, Google Cloud Healthcare, to discuss how these technological advancements are affecting medical groups and cut through what is conjecture to find the real-world applications of these promising technologies. Here are nine key takeaways from the session: 1. A.I. is old, but its biggest uses are new A.I. was originally developed in the 1950s and 1960s and enjoyed a lot of hype in the 1970s, Muelly said, but those initial neural networks required bigger data sets and more computing power that didn’t exist until the mid-2000s. Bigger data sets — such as the Stanford Vision Lab, which trained neural networks to catalog and annotate more than 14 million images2 — open a wider range of potential applications. “Since around 2011 or so, any product that you see from any of the larger tech companies has some form or shape of machine learning, A.I. or, more specifically, deep learning, built in,” Muelly said. 2. Data quantity and quality make a big difference Finding the right fit for using A.I. and machine learning matters a lot, according to Cuesta. “It seems like A.I. and machine learning can sort of figure everything out,” Cuesta said. “Turns out that’s not right. … Machine intelligence is only as smart as the data that it trained on.” Pointing to the Stanford example, Cuesta said that good data that’s curated well can make a very intelligent machine. But “if you have less data, or the data is of poor quality, or is not tagged well, you’re going to end up with a machine intelligence that is not very smart,” Cuesta cautioned, “and it will figure out bad answers very fast.” Muelly added that the models used in A.I. and machine learning boil down to “beating the pattern of the data into them” — showing multiple examples for the machine to learn to easily spot variances. “It doesn’t necessarily have to be something that it has seen before, but it has to be pretty close to that,” Muelly said. “That is the way that these are trained.” 3. There are two major practical applications of A.I. today According to Siegel, the two most practical applications of A.I. in the medical care setting today are classification and anomaly detection. Pointing to Muelly’s note about pattern recognition, Siegel said that recognition is important in creating structured data through classifications, such as if a practice receives countless pages of patient charts and other clinical documentation via fax. “Patients’ full charts that just get faxed in and get brought into the EHR as an electronic attachment — each one of those pages could represent all different sorts of structured data,” Siegel said. Some algorithms have been developed to help classify what types of data those pages contain, so as to classify them as referral letters, historical notes or lab panels, for example. The anomaly detection aspect of A.I. often is used by payers, Siegel added, wherein the machine intelligence is trained to spot if the claim being submitted by a practice looks different from a similar service that normally would be billed. But in any case, “the machine intelligence only knows what it’s trained on,” Siegel cautioned. “If you have an algorithm that’s trained on inpatient representative data, it’s of some usefulness but not going to be as entirely applicable in the ambulatory setting.” 4. Clinical uses have come along quickly despite more work, high stakes The clinical uses of A.I. have established themselves and proven valuable “both from a quality perspective and from a productivity perspective,” Muelly said, despite the clear concerns about the significant impact they might have on patient care. That value has pushed many clinical A.I. models along despite requiring a lot more validation and numerous regulatory mandates. Conversely, a number of nonclinical applications of A.I. are relatively simple to build and “the stakes are not quite the same,” Muelly said, pointing to the example of A.I. techniques to automatically code a claim. When the A.I. model doesn’t perform correctly in that setting, “your claim may get rejected because it was incorrectly coded, but [there was] no ultimate impact on patient care,” Muelly said. “You can rectify that; if you misclassify a lung nodule that turns out to be cancer, [that’s] much higher risk.” 5. Finding common structure in data is essential The amount of data that healthcare organizations collect and are responsible for managing has increased exponentially in the past decade, Siegel said, so the opportunity to use that data for machine learning “means that the responsibility in how we’re collecting that data, tagging that data, managing it, making sure that we’re keeping track of all of it — that starts now.” That represents a “master data management problem” today for medical practices, Siegel said. Good data hygiene is crucial: “You serve your group best” by ensuring clean, accurate data before launching an A.I. or machine learning initiative, Siegel added. Muelly noted that in his work at Google, “we typically spend about 80% of the time actually on data quality. … The vast majority is really ensuring that you have good quality data.” Muelly said that if you believe A.I. applications will be relevant in your practice, your next best step is ensuring that your data quality is up to a certain standard. In the ambulatory clinic space, that means managing a variety of data formats from lab vendors, payers and others that require some degree of standardization. “You have to pick the code system that you want to standardize on,” Siegel said. “Less is more to a certain extent. If you can have a base set of standard data elements that you can collect every time and guarantee them with high quality, that’s a place to start.” The FHIR (Fast Healthcare Interoperability Resources) standard for application programming interfaces (APIs), as a highly touted means for data formatting, represents an opportunity for early adoption that could help practices as the healthcare industry pushes toward interoperability, Siegel added. 6. Despite the hype, you need to set realistic expectations Muelly, who also works as a practicing radiologist in his time outside of Google, said that change management is a challenge for data standardization work. “There can be a resistance,” Muelly said, pointing to a shift in the past 10 to 15 years from free-text reports — largely dictated by radiologists — to more structured reporting. “You can’t go from quality data in EHRs, which generally is not that great, to perfect data for training [machine learning] in a day,” Muelly noted. “It really requires that shift over time.” Siegel said that he has seen practice leaders grapple with “massive disappointment” thinking that a decade of electronic note documentation is sufficient to dive into an A.I. project, only to learn that more work needs to go into making the data useful for an A.I. application. Instead, it’s best to think of A.I. as a “technology sandwich,” in which A.I. is placed in the middle of a workflow as a rules engine: As data is inputted into the system, the work done by nurses to tag classifications onto data instead can shift to the A.I. — allowing those nurses to return to clinical activity and prioritizing what the A.I. has sorted to allow the human element to be the final component. “I think that’s where we’re going to find the most satisfaction from employees and practitioners adopting A.I. in the practice,” Siegel said. “It should be a help to us. It doesn’t necessarily have the capability to replace human work.” 7. Understand the value of a “human in the loop” Muelly pointed to his work at Google as an example of where having a “human in the loop” of an A.I. model is vital. Something as basic as the spam filter within the Gmail platform learns from the feedback it receives from users. Providing feedback to A.I. models in healthcare gives similar information that helps improve the machine learning. “In the clinical setting, you are not going to completely remove the human,” Muelly said, but noted that some useful applications of A.I. in which human user feedback will be helpful might include presorting messages from patients or routing messages to the right team member, whether it’s the physician, nurses or front desk staff. “All that can be very easily automated and ultimately still require the human in the loop, which also serves as a fail-safe if one of the classifications goes wrong,” Muelly said. 8. Doctors should not worry about being replaced by A.I. Muelly described his work life as 90% at Google and the remaining 10% as a clinician. “The reason I do that is not because I believe A.I. will replace my job as a radiologist,” Muelly noted. When it comes to clinician productivity and many elements of practice management, Muelly said he believes there is “tremendous opportunity” for A.I. applications, but it takes years for them to make a real impact — and it won’t pose an existential threat to the medical profession when it does. 9. Free text isn’t going away, but it will improve As practices look to document more care elements for quality measures properly and effectively, many clinicians are hopeful that natural language abstraction or processing will help in analyzing free text documented in patient records and will lead to better claims and quality reporting. Muelly said that drop-down menus in EHRs and taught phrases are still going to be needed in balance with the promise of natural language processing to make better sense of free text in records. “Technology is advancing every day. Natural language processing has made huge leaps,” Muelly said. “A lot can be done, but it continues to be more error-prone than human analysis.” Notes: IDC. Worldwide Semiannual Artificial Intelligence Systems Spending Guide. March 11, 2019. Available from: bit.ly/2qCExol. Abate T. “Stanford team creates computer vision algorithm that can describe photos.” Stanford Engineering. Nov. 18, 2017. Available from: stanford.io/2QVGp68.