Leo Celi and the 'Holy Grail of Personalized Medicine'

Nearly everyone extols the need to bust data silos. Meet the man who’s doing it.

Leo Celi, MD, is striving to advance precision medicine. Can he help transform healthcare?

AN AWAKENING CAN COME IN A FLASH, or it can occur over years, as dogma gradually breaks down in the face of data. For Leo Anthony Celi, MD, there was never a “eureka moment,” as he puts it, during his journey from studying medicine in the Philippines to becoming clinical research director at the Massachusetts Institute of Technology (MIT) Laboratory of Computational Physiology. Instead, it took rigorous training, along with extensive experience in clinical practice, before he recognized that the art of medicine can be woefully unscientific.

Celi, 51, says that realization can be “very disturbing.” He could have found it demoralizing too. Yet colleagues describe an investigator, professor, and physician with seemingly boundless enthusiasm. “Frankly, I don’t think I know anyone who is as energetic as he is,” says Peter Szolovits, PhD, who has worked at MIT for more than 30 years and is now head of its Clinical Decision Making Group, where he collaborates frequently with Celi. “I suspect he’ll continue to work like that when he’s 80.”

It might take that long before Celi’s vision for healthcare is even partially realized. The problem with the art of medicine is that it varies from doctor to doctor. As Szolovits explains, “What fraction of the clinical decisions you’ll make as a practicing doctor have had randomized clinical trials? If you’re an optimist, it’s 10% to 15%.” After more than 2 decades working in the intensive care unit (ICU), Celi knows this firsthand. “When it comes to what should I do for this 86-year-old gentleman in front of me with all these chronic conditions, it’s all guessing because there was no experiment that only enrolled 86-year-old men with his conditions,” he says. “That’s why we see so much variation in care.”

>>Read: Hunting for the Heart of a Changing Community

“How can we make the practice of medicine more transparent and less likely to be affected by bias?” Celi asks. “To me, the only way to do that is by partnering with computers that can process a voluminous amount of data so much better than humans can.”

Virtually every industry is integrating data science, and practitioners in those fields often feel as though their expertise is under attack. It’s compelling to hear Celi speak on the fallibility of physicians in light of his credentials: Along with master’s degrees in public health from Harvard University and biomedical informatics from MIT, he completed specialized residency fellowships at the Cleveland Clinic, Stanford University, and Harvard University. “People are always paranoid that it’s going to be doctors versus computers,” Celi says. “If I were a patient, I would go to a doctor with a computer.”

Celi and his team foresee an optimal version of doctors with computers, dynamic clinical data mining. They imagine a system in which the “collective experience” of every possible patient on earth is aggregated and analyzed, as they’ve written. “This approach would interrogate data to suggest next-step options and weigh the risks and benefits of a treatment or test for a specific patient.” It would be, in short, “the Holy Grail of personalized medicine.”

That concept is made possible by electronic medical records (EMRs) and made impossible, for now, by the widespread reluctance to share patient data. One remarkable exception is being modeled by the lab at MIT where Celi works, using data from the ICU at Beth Israel Deaconess Medical Center in Boston, where he practices. The lab compiled elaborate data from more than 50,000 hospital admissions and has made the database accessible, for free, to thousands of investigators and students around the world. Granted, nobody can form conclusive insights by studying a decade’s worth of patient data from 1 unit at 1 hospital, but the database has been an amazing resource for exploring research questions and—more important—demonstrating how big data and machine learning could radically augment the wisdom behind treatment decisions.

From residents at Beth Israel Deaconess to professors at MIT, people who know Celi rave about his ability to facilitate collaboration among clinicians, data scientists, engineers, statisticians, and whoever else is needed for a project to succeed. In the ICU, he learned how the wellbeing of patients depends on a “collaborative culture,” he says. That’s also the only way forward to healthcare’s holy grail.

Melding Medicine and Computing

The same year Leo Celi was born in the Philippines, psychologists at the Oregon Research Institute were performing groundbreaking experiments on the shortcomings of human judgment. As it happens, the first of their tests observed medical professionals. When examining stomach x-rays for cancer, radiologists “tended to describe their thought processes as subtle and complex and difficult to model,” as Michael Lewis recounts in his latest book, The Undoing Project. Yet computers using a simple algorithm did better at diagnosing cancer than did the doctors being studied. For the lead Oregon investigator, the results were “generally terrifying.”

In college and then medical school at the University of the Philippines in Manila, Celi certainly had no encounters with data science. His mother was a nurse, and his sisters and cousins also became nurses—in their country, a career in medicine was a guaranteed way out of poverty. When Celi entered medical school, he found “it was all memorization and embracing concepts that are handed to you.” His professors imparted what he later came to see as a naïve sense of confidence.

“In school examinations, there are always multiple choice and true or false, when in fact there are no correct answers,” Celi explains. “It depends on what patients value more: Do they value this probability of improvement from an intervention versus this probability of harm from a side effect? As doctors, we can’t help them navigate that because we ourselves are very uncomfortable with probabilities.”

Now Celi believes medical schools and residency programs “need to be transformed radically,” with data science incorporated into core curricula so that doctors can make better use of EMRs and feel comfortable working with investigators. (New York University School of Medicine is leading the way on this.) He and his lab partners at MIT outlined that recommendation in their 2016 paper Bridging the Health Data Divide. Clinicians are inclined to believe they know what’s best for patients, while engineers can find doctors stubbornly set in their ways. Data silos are notorious in healthcare, but expertise has been placed in silos, too.

After earning his medical degree in 1990, Celi followed the rest of his family to the United States, where he received training in internal medicine at Cleveland Clinic, in infectious diseases at Harvard, and then in critical care medicine at Stanford. In 1999, he joined a startup out of Johns Hopkins University called Visicu, where he worked with software and hardware engineers to design technology that would improve the efficiency of ICUs.

Around that time, Roger Mark, MD, PhD, was pushing a computer on a cart through the ICU at Beth Israel Deaconess Medical Center, plugging it into a monitor to record data with the permission of 1 patient at a time. In an effort to study arrhythmias, Mark recalls, “we’d ask the nurse, ‘Who had a bad night? Who’s unstable?’ Then we’d start recording, and of course the patient would be stable as a rock that day.”

Mark directs the Laboratory of Computational Physiology at MIT, and this primitive method of data collection was the origin of Medical Information Mart for Intensive Care, or MIMIC, the database with de-identified records from patients admitted to the hospital’s ICU from 2001 to 2012. (It was made possible after the institutional review board at Beth Israel approved of Mark’s gathering data without individual consent.) Mark recognized that each patient’s experience was, in essence, an experiment. Years later, people like Celi would be able to parse that data for all sorts of unanticipated findings.

In 2002, Celi accepted a faculty position at the University of Otago in New Zealand, where he developed safety protocols for the ICU while continuing to practice in the emergency department. A medical student there, Xaviour Walker, was struck by Celi’s “thirst for knowledge” and his example as a “hugely generous and humble person. He always talked to nurses, techs, students, and administration staff to get to know them as people.” Celi encouraged Walker to continue studying in the United States. By 2007, Celi had decided that he, too, needed additional education—the art of medicine, as he’d been taught it, was failing patients. So he returned to Cambridge, Massachusetts.

As Celi began dual master’s programs at MIT and Harvard, the National Institutes of Health awarded Roger Mark’s lab a major grant to expand MIMIC. “I went back to school to try to improve the way we deliver care, and then suddenly the government granted funding to build the tools I was interested in building,” Celi says. “It was almost too good to be true.”

For his master’s thesis at MIT (with Peter Szolovits’ supervision), Celi used MIMIC to analyze 1400 patients admitted to the ICU with acute kidney injury and found that traditional models for predicting hospital mortality, which use prospective observational studies, were less effective than customized modeling would be using retrospective data. In conclusion, he noted that “the question remains whether clinicians will embrace this approach. Will we be able to convince them that information from a very large cohort of patients whose clinical course is stored in an electronic database might be more reliable than a composite of the patients they have encountered in the past whose clinical course may be imperfectly stored in their memory?”

Nearly a decade later, many clinicians remain unconvinced. Szolovits, for his part, maintains hope that the medical community will embrace data science. But, he adds, “I was optimistic in 1974, and it’s taken a while. I would take my optimism with a grain of salt.”

The Physician and the Algorithm

These days, at any given moment, Celi is involved in about 25 research projects. He continues to work as an internist in the Beth Israel Deaconess ICU for a week each month—as an investigator, he considers it crucial to remain in touch with the challenges facing clinicians. When he’s not at the hospital, Celi typically starts his day at 7 AM on conference calls with collaborators based all over the world. He works at his office on MIT’s campus from 9 AM to 6 PM and then returns home for more conference calls until 11 PM or so. Somehow he finds an hour every day for kickboxing. “Wherever we travel in the world,” explains one of his lab partners and kickboxing converts, Matthieu Komorowski, MD, “we bring the boxing gloves and kicking shields.”

On Friday mornings, Celi teaches a course in the Health Science and Technology program run jointly by MIT and Harvard Medical School. The course, Global Health Informatics, allows students to conduct research using MIMIC. The database includes, for example, progress notes from care providers; timestamped physiological measurements; continuous intravenous drip medications and fluid balances; laboratory test results; and data from the hospital’s 2 critical care information systems, Philips CareVue and iMDsoft MetaVision. All this data collection was possible without disturbing ICU workflow. Mark estimates that 6000 people around the world have completed the credentialing process to use MIMIC.

Kenneth Mukamal, MD, has worked at Beth Israel Deaconess since 1994 and was skeptical when Celi began using natural language processing to analyze doctors’ notes to find out what people were taking before they came to the hospital. Mukamal assigned a resident to go back and compare 100 charts to Celi’s method. “I’m not embarrassed to say that the algorithm was almost perfect. Leo was right, and I was wrong.”

Celi’s lab was designing “really ingenious ways of using what most people like me didn’t even consider data,” Mukamal says. With the natural language processing algorithm in hand, the team analyzed 11,490 patients in 2013 and documented how diuretics can dangerously lower magnesium levels. “That work had a direct impact on how physicians across the United States prescribe these pills,” Mukamal says. “I’m not sure even Leo has a sense of how important some of those papers were.”

At the beginning of the semester, Celi’s course invites anyone at the hospital to pitch ideas that the students will help investigate. Celi is like a “magnet,” as Mark puts it, for attracting diverse participation. For instance, 2 years ago a resident at Beth Israel Deaconess, Robert Stretch, MD, went to Celi with the idea of studying the effect of boarding—when patients are assigned to open beds in a different section of the ICU—on mortality. “Leo set me up with a machine-learning expert who also happened to have expertise in some sophisticated statistical methods,” Stretch says. “When I agreed that the statistical methodology Leo recommended made sense, he said, ‘Great, I’ve got exactly the right person for you,’ and within 48 hours we had the core of a research group for a publication.” Their research did, in fact, indicate that boarding critically ill patients increased mortality.

Marzyeh Ghassemi, PhD, was working on machine-learning algorithms in Peter Szolovits’ lab when Celi went to her with a research question: How do selective serotonin reuptake inhibitors affect in-hospital mortality? “When you have a collaboration between a clinician who has hunches about treatment, and a machine-learning person who wants to look at data in specific ways, that’s when you learn things more robustly,” Ghassemi says.

To draw meaningful insights, investigators will need access to data from tens of millions of patients, a lesson learned from genomics. At events where investigators explore MIMIC, Celi reminds participants that their findings “might be completely useless.”

One of his priorities right now is persuading hospitals around the world to build their own MIMICs, and not just for the ICU. “My main focus was just [EMRs],” Celi says, “but we’re realizing we need to have a comprehensive representation of the patient,” even beyond clinical data. As long as vendors see data as a commodity worth hoarding, or are overly paranoid about patient privacy, data science will remain stalled. Celi says federal regulation is necessary to compel sharing. “To me,” he explains, “the data really belong to the patients and the public.”

In 2007, Philips Healthcare acquired Visicu, the startup where Celi once worked, for $430 million. Using their connections, he and Mark were able to persuade Philips to release a 200,000-patient subset of its data archive for public use. This “eICU,” like MIMIC, is accessible to investigators and students for free.

The dearth of data has limited the ability to test new technology, “and that’s why people are still very skeptical of the true value of machine learning,” Celi says. Most of the progress with artificial intelligence in healthcare has involved image recognition, for which there are clean data sets available. “I could argue that those are low-hanging fruits,” Celi says. “Improved diagnosis is obviously valuable, but improved treatment is going to benefit so much more from machine learning.”

He says is, not could, but there’s no indication, yet, of when.

A Bright Future for Big Data?

While pursuing dual master’s degrees and continuing to practice on weekends in the ICU, Celi and another master’s student, Kenneth Paik, MD, launched a program called Sana, which brought together students in medicine, engineering, and business to improve healthcare in developing countries using new technology. Sana (which means “healthy” or “life” in several languages) had considerable success early on, but the challenge became how to sustain those efforts. Celi and his partners realized their greatest objective needed to be supporting people in those countries to build their own health information systems.

Those who believe in the urgent importance of data sharing can exude an evangelical’s passion, yet they continue to struggle to win over enough converts. Szolovits described a recent meeting with the head of Children’s Hospital Los Angeles. Over dinner, the administrator agreed to contribute his hospital’s data to MIMIC, according to Szolovits, which would have tripled MIMIC’s pediatric data. “But then he sobered up the next day and didn’t do it.”

“Everyone agrees this is the right thing to do,” Szolovits says, “but then they find a reason not to.”

Sana is now under the umbrella of MIT Critical Data, and Celi and a small team travel almost every month to host events—over the past 4 years, they’ve been to the United Kingdom, Ireland, Germany, Italy, France, Greece, Spain, Brazil, Argentina, Colombia, Mexico, Uganda, China, Taiwan, Thailand, Singapore, Australia, and the Philippines. “The best part of this job is meeting all these brilliant students from around the world,” Celi says. His team partners with local institutions to organize “datathons,” often using MIMIC, which aim to grow excitement around data science.

“Young people are more collaborative in the way they approach problem solving,” he says. “They will transform how we do things. If it weren’t for them, we probably would have burned out a long time ago.”

Danny Funt is a freelance journalist based in New York. He can be reached on Twitter @dannyfunt.

Related

Rise of the Anti-Opioid Algorithm

An Innovative Way of Collecting DNA Samples Should Have Researchers Salivating

Interoperability During Disasters: Lessons from Tragedy