DataScience@NIH Blog

Today’s Interim Associate Director for Data Science and a 19^th century Surgeon General: They served in different eras under different conditions, but the similarities between the challenges they faced are striking.

A Modern Leader

Last summer Francis Collins, MD, Director of the National Institutes of Health (NIH), selected Patricia Flatley Brennan, RN, PhD, as the new Director of the National Library of Medicine (NLM). In January of this year Dr. Brennan also assumed the role of NIH Interim Associate Director for Data Science (ADDS).

Dr. Brennan faces a daunting task in these dual roles. Data and information are being generated on a scale never seen before, especially in the health sciences. Moreover, it is being generated in new ways, with new technologies, and often only captured in ephemeral media like web-based discussion forums, databases, and even blogs like this one. Gathering, organizing, preserving, and providing access to the salient aspects of this new knowledge is just as difficult as it is important. An unprecedented challenge, to be sure… but not one without historical parallels.

Lessons from History

The historian James A. Huston, discussing the value of history, wrote that "History provides perspective for the judgment of ideas, policies, and procedures… a much more important contribution of history than its direct lessons is the pool of vicarious experience which it provides—experience which is the raw material of imagination.” Huston elaborated: “It is a function of history to provide rich experience out of which imaginative leaders will create new methods to meet new situations."[1]

While the challenges of today’s data-driven world are thoroughly modern and unique, there are historical analogies that can provide helpful perspectives. One hundred and fifty-five years ago, the Surgeon General, who in some ways was Dr. Brennan's NLM predecessor, faced a similar challenge.

A Time for Bold Leadership

William A. Hammond was the newly-appointed Surgeon General of the US Army as the country entered its second year of civil war in 1862. His office included a tiny library of just a few hundred volumes—a library which would eventually grow into today's National Library of Medicine. Hammond represented a significant change in the leadership of his organization. The fourth man to hold the title "Surgeon General" in the medical department that had been reorganized six decades earlier, Hammond was the first to be selected based purely on merit rather than simple seniority. The Secretary of War could easily have continued the tradition of appointing the highest-ranking and longest-serving officer in the medical department as Surgeon General, but he was looking for something different. The country was at war, the challenges ahead were immense, and he needed someone who could bring fresh vision, audacious ideas, and bold leadership to the position. He saw those traits in the young William Hammond.

A New Age

In 2016, NIH Director Francis Collins was looking for a new NLM Director. As with Hammond’s appointment, this would be the fourth person to hold the Director title, 60 years after the Library had been reorganized as a civilian institution with its current name. Collins needed someone who could lead the National Library of Medicine into the age of data science, someone who could address a new kind of challenge. Dr. Collins selected Dr. Brennan to meet this challenge. Like Hammond, she was an unorthodox choice—the first nurse and the first non-MD to serve as the NLM Director. There are substantial parallels between Brennan and Hammond and the challenges their organizations faced. The challenges and opportunities we see at NLM today make it worthwhile to examine these parallels.

Examining the Parallels

In 1862, Hammond recognized that the world of medical knowledge was changing. This change had been emerging for decades, as the scientific revolution prompted ingenious and resourceful innovators like Edward Jenner, René Laennec, and William Beaumont to develop new medical techniques and procedures. Hammond could see that such breakthroughs and advances would no longer be rare, isolated incidents. Doctors and surgeons would be seeing new types of injuries on a vast scale, and they would inevitably, necessarily devise new techniques and develop new knowledge from these experiences.

Abundant New Information

Hammond also recognized that, while there was value in the individual experiences and knowledge gained by busy physicians in the field, if he could harness and share this abundant new information he could make it exponentially more effective across the Army Medical Department and beyond. Nothing like this had been done before. Not only was the scope of knowledge to be captured unprecedented, but the effort to proactively acquire and preserve it, to have systems in place before the information had been generated, was prescient and novel.

A Focal Point for Medical Knowledge

Hammond approached this challenge the way he approached everything: with methodical, deliberate action, foresight, and strategy. He found the best people to help him—imaginative, enterprising medical officers like George A. Otis, Joseph J. Woodward, and John H. Brinton—and he gave them the necessary resources to develop their expertise and fulfil his vision. Most importantly, he firmly established a vision for the future of the Surgeon General's library as a focal point for the accumulation and distribution of medical knowledge.

Hammond directed surgeons to send in specimens, much as today's researchers have been encouraged to share their findings through a range of health services research and public health information programs. John Brinton and George Otis received barrels of surgical specimens from the field; Dr. Brennan's team is receiving "barrels" of digital data in much the same way. Hammond collected physical objects, like new medical instruments and improvised surgical devices, for the information they could convey. Today the Internet of Things—ubiquitous interconnected devices and sensors—allows objects to generate their own data, conveying information in new ways.

A new information medium—photography—also provided exciting new possibilities for Hammond and his field officers in 1862, but it required new ways of analyzing and sharing information. Photographs and photographic plates could yield new insights and the obvious advantages of a visual record, but they were fragile, easily corrupted, and hard to reproduce and disseminate. Today we have access to data sets full of potential insights and opportunities for discovery, but they can be just as hard to preserve and share as those early photos.

Deriving the Greatest Benefit

As we look out over the information landscape in 2017 we see the emerging changes, just as Hammond did in 1862. Powerful computers are inexpensive and widely available. Cloud-based networks enable global collaboration. As in the Civil War, the walls of the laboratory are disappearing and knowledge generation is now possible for any creative and enterprising researcher. And, as in the Civil War, this unprecedented flow of information needs to be captured and preserved to derive the greatest benefit. This is the challenge facing Dr. Brennan and the NLM, and it is every bit as daunting as the challenge Hammond faced. Fortunately, Dr. Brennan and her team appreciate that history can provide the vicarious experience, the "raw material of imagination," necessary to succeed.

-----

About the Author:

Kenneth M. Koyle is the Deputy Chief of the History of Medicine Division at the National Library of Medicine, whose mission is to collect, preserve, interpret, and provide access to one of the world’s richest collections of historical material related to human health and disease. Ken’s primary area of historical expertise is 19th century military medicine.

The National Library of Medicine supports discovery. We help investigators disseminate their findings, and we help others see what has been done and build on those results. And, analyzing the library collection itself—the journal articles on a topic or the sequences in our gene databases—can yield its own discoveries as researchers look beyond the specifics to discern patterns or trends in the whole.

In this era of data-driven discovery, however, we need to do more.

First, we need a PubMed-equivalent for data. While the name “PubData” doesn’t do it for me, we need those basic functions.

After all, data-driven discovery begins with discovering the data. We must make that as easy as possible.

Just as PubMed provides citations to articles in selected journals, we need to compile a common catalog of biomedical data sets. And, as the NLM Literature Selection Technical Review Committee brings together experts to review journals and assess their quality to identify what should be indexed in PubMed, we need an experienced group to identify and select data sets based on predetermined criteria. Those criteria, in turn, will help set the standard for quality data.

Some PubMed citations link to the full-text article. That full-text might be in PubMed Central—which means the article is stored here—but most are not, so we simply link to them where they are, whether that’s a publisher’s site or another library’s repository. Similarly, while some data sets might be deposited in NIH-hosted repositories, most won’t need to be. We can link to them instead.

PubMed offers other value-added services—from standardized metadata to suggestions of similar articles to related data in other databases. The data side of the house will need similar built-in services: links to articles that used a data set, explanations of how a data set was enriched or modified, metadata that fully describe the origins and content of the data.

But, that is only the beginning.

Data-driven discovery also relies upon proven methods to investigate the data, create predictive or explanatory models, and conduct a range of operations, so we need to think about how to create a library of models, both those based on statistics and those drawn from operations research and optimization.

We’re just beginning to consider what a library of models might look like. Many of the services we use for curating the literature, such as a common vocabulary and a standard way of attaching metadata, will be necessary. We will want to know the class of modeling tools and a model’s provenance and intended use. It will also likely be useful to tie a model to the data sets investigated with it and to the articles derived from that model’s use.

I’m sure there’s much more. What else would you include?

We’re at the beginning. What do we need to make this work?

Around NIH we’re getting ready for Pi Day, an annual celebration of the irrational number Pi, 3.14159.

Yes, we had planned to celebrate Pi Day at its appointed time—March 14 (a/k/a 3.14)—but as sometimes happens when baking real pies, the unexpected occurred—in this case, a freak mid-March snowstorm. Between the threat of a blizzard, the reality of six inches of snow, and the key speaker’s family emergency, we had to postpone, but we are the largest biomedical science operation in the world—we can declare another day Pi Day!

And so we did.

For this year only (we hope), Pi Day is 5.18, not because of any numerical significance, but because that’s when we could find the space, the place, and the opportunity to celebrate.

We’ll mark the occasion with a day of events and activities celebrating the intersection between the quantitative and biomedical sciences. Most events will take place at the NIH Clinical Center (Building 10), but if you are unable to join us in person, you can still view the main lecture and lightning talks live on NIH Videocast or engage with us on Twitter using #nih_piday.

Bonnie Berger, Simons Professor of Mathematics at MIT with a joint appointment in Computer Science, will deliver the main lecture, "Mathematics of Biomedical Data Science," from 1-2 pm in the Masur Auditorium (Building 10).

Earlier that day (11am-12pm, Masur), ten data scientists from across NIH will present lightning talks using thee slides to convey one idea in four minutes. (3+1+4. 3.14. Get it?) Tours of the NIH Data Center and the Pi Day Networking Event, featuring posters, demonstrations, exhibit tables, and pie (!), will bookend the lightning talks. Following Berger’s lecture, the day will end with a special hands-on workshop on reproducibility in science using Python and GitHub.

Pi Day events are free and open to the public, though the data center tours and the reproducibility workshop both require advance registration. (Registration: tour | workshop)

So mark your calendars and come celebrate Pi Day with us! I look forward to seeing you there.

Data scientists in the health sciences work primarily with data generated during experiments, from specialized machines such as next-generation gene sequencing devices or mass spectrometers. But there’s a data type becoming increasingly important to discovery in health care: patient-generated data.

Patient-generated data arises from observations that can be made only by the individual, often during the course of everyday living. Most often these data are self-reported, using a rating scale that solicits the individual’s perceptions, for example, of pain levels or sleep quality. Devices can also be used to acquire patient-generated data, such as a glucometer that captures the concentration of glucose (sugar) in the blood. I consider these “professionally-defined, patient-provided” data.

There’s another type of patient-generated data, one I’ll call “patient-defined, patient-provided.” One example comes from my friend’s mother. She calls one specific data element “wooziness,” and it refers to the feeling she gets when she’s dehydrated. Her thinking slows, and she has a sense of discomfort.

It’s not professionally defined, but “wooziness” provides her with important health information. It heralds the need to act, in her case, to drink water. She can talk about feeling more or less woozy; she knows which days she felt woozy; and she’s beginning to understand what causes it and what mitigates it.

Such patient-generated data do not follow the rules of formal terminologies—these data might not even map to specific anatomical or physiological phenomena—but they tell us something about the person and his/her health experiences.

And that makes patient-generated data powerful.

They give us a window into the person’s everyday health experience. They provide a valuable complement to professionally obtained data, such as clinical observations or biological samples, and they form the basis of assessing patient-reported outcomes, an increasingly accepted indicator of the quality of care.

And since one significant goal of data science is to better characterize the context of health, I believe increasingly incorporating patient-generated data can help us do that.

Like everything in the realm of data science, that will come with its challenges, but for now, if the data you’re working with were contributed by a person, give a silent word of thanks to your invisible partner in discovery. I know I do!

Those of you born before 1980 or the inveterate fans of 1960s science fiction know what tribbles are. These small, soft, furry, purring space aliens appeared on a memorable 1967 episode of “Star Trek.”

The pleasure and comfort they bring to humans come with a significant downside: tribbles do nothing but eat and reproduce. Born pregnant and with no known predators, tribbles multiply exponentially, quickly consuming anything edible. Their care and feeding exhaust all available resources, leaving little reserve for other pursuits and other species.

Sounds a bit like research-generated data, doesn’t it?

Not long ago (but long after the first tribbles appeared on “Star Trek”) a research project produced a resolution to the hypothesis that initiated the project. Questions were posed, hypotheses generated, data acquired, analytics applied, and interpretation emerged. Done and done. But since the late 1990s or so, many research projects have also generated data sets that promise—or at least hint at—future discoveries.

Researchers the world over spend substantial funds curating and storing these data sets and ensuring they are FAIR (findable, accessible, interoperable, and re-usable). NIH alone spends hundreds of millions of dollars annually to ensure high-value data sets are securely stored and made available for future use and to foster among investigators an appreciation for and willingness to engage in data-driven discovery.   

All of which leads me to conclude that data science can learn a lot from “Star Trek” and its tribble troubles.

Like tribbles, data are attractive and pleasing to many. Some species, like Klingons, abhor tribbles, and I suspect a few out there abhor data or see data science as another scientific fad. Just as tribbles and their perpetual offspring can place extreme demands on a system, so too can data sets, especially those left to grow without purpose or design that end up competing for scarce research dollars. (Fortunately, we’re not yet at the point of having to choose between new investigations and data-driven science, but that day may come.) And the solution to the Enterprise’s tribble infestation—transporting them to a nearby Klingon vessel—sounds a bit like the hope that “the cloud” will solve our data science challenges, when such a move solves only one problem and shifts the others to a new environment.

Despite those challenges, some data science zealots advocate replacing experimental science with data-driven discovery, but I recommend a more balanced approach. As the Federation’s Prime Directive makes clear, every species and society should be allowed to follow its normal cultural evolution, and, to me, data science is part of the evolution of scholarly discovery.

I invite you to partner with me and the NIH to ensure a principled approach to data science.  

What steps should NIH take to make sure the “data tribbles” don’t crowd out the full range of discovery strategies needed to deliver the greatest benefit to society?