Predictive Medicine: Where the Pentagon and Silicon Valley Could Build a Bridge in Artificial Intelligence

With over 9.4 million patients in the Department of Defense (DoD) healthcare system and another 9 million patients enrolled in the Department of Veterans Affairs (VA) healthcare system, the federal government has the largest healthcare system in the world. In the era of machine learning, this translates to the most comprehensive healthcare dataset in the world. The vastness of the DoD’s dataset combined with the department’s commitment to basic biological surveillance yields a unique opportunity to create the best artificial intelligence–driven healthcare system in the world. The goal of predictive medicine is urgent and achievable, and would have transformative effects throughout the medical profession.

Why should the DoD care? Talking cancer alone, $1.7 billion of the Defense Health Agency’s annual budget is spent on cancer, and that fraction is growing. Cancer requires the most expensive type of care, and the patient population continues to grow. The medical profession still does not know all the causes for the various types of cancer. From a readiness perspective, cancer is responsible for many medical separations, lost duty time (for those who can return to duty), and loss of livelihood and normalcy for dependents. Some patients requiring radiation and chemotherapy are not always lucky enough to survive the long process, and those that do have a potential for long term health complications. Moreover, diagnostic errors are the most common cause of malpractice, resulting in large Defense Health Agency (DHA) payouts.

More broadly, from 2001 to 2016, funding for central medical programs increased by 84 percent and is the single fastest growing cost of DoD support functions. The cost of providing that care has increased rapidly as a share of the defense budget over the past decade, out-pacing growth in the economy, growth in per capita health care spending in the United States, and growth in funding for DoD’s base budget. The United States is equally falling behind nations like China and Canada, where data sharing is easier, in artificial intelligence (AI) applications for healthcare.

Predictive medicine can address force readiness challenges. Given all the data, not only can DHA physicians and researchers make more accurate predictions, they can discover relatively strong links between unknown baskets of variables. For example, there may be combinations of genomics and environmental exposures that make certain warfighters more vulnerable to injury or infections.

Medical innovation is nothing new to the Department of Defense. Following each conflict, battlefield casualty response of medics and corpsmen has laid the groundwork for civilian first responder training. DoD created GPS, global communication satellites, the internet, and many other components of the global telecommunication infrastructure, which laid the groundwork for Telemedicine. Vaccine research in Smallpox, Yellow Fever, and Adenovirus all started in the DoD. Becoming the leader in AI-based medicine would be another step in a long history of DoD-led innovation, with spillover effects in the civilian world.

There are many well-known challenges to implementing machine learning and AI in healthcare. The first is the lack of curated data sets, which are required to train AI via human-supervised learning. Robust curated data sets that have both the breadth and depth for training are essential, but frequently hard to access. This is where the DoD has a clear advantage. Required examinations for service members include in-processing physicals, care while on active duty, and a final separation examination. Life after retirement continues with treatments in the VA. Baseline data, regular biological tests, injuries, regular annual exams, and sick call visits, are thoroughly documented and stored in the Defense Department medical record system.

By the numbers, the Military Health System has 6 theater hospitals, 57 medical centers, 364 ambulatory clinics, 117 naval ships, 281 dental clinics, 255 veterinary facilities, 2 hospital ships, and 17 submarines that generate medical data. A single day in the Military Health System represents 1.5 million International Classification of Disease (ICD) 10 codes, approximately 87,000 prescription orders, approximately 54,000 clinical notes, approximately 13,000 radiology orders, 160,000 output encounters, approximately 10,000 telephone consults, and approximately 64,000 lab orders. During the last 10 years, the Military Health System has stockpiled data on 5.4 billion ICD9/ICD10 diagnoses, 2.8 billion procedures, 442 million outpatient encounters, 202 million prescriptions, 234 million lab orders, 145 million clinical notes, 47 million radiology scans, 29 million telephone consults, and 2.5 million brain MRIs. This represents an unprecedented and comprehensive longitudinal clinical history.

In addition, the DoD also has unique data sets that no other single place in the global medical community maintains. For example, the Joint Pathology Center (JPC) is the largest repository of cancer tissue in human history. Over a century of tissue samples and diagnostic data have been gathered by pathologists from around the world. As of 2008, the JPC has amassed 7.4 million cases, 32 million tissue blocks, and 55 million slides. Each slide represents roughly 5 gigabytes of data, for a total of 275 petabytes of labeled images. The tissue blocks from each case contain terabytes of additional information genetic and proteomic information. For many years the JPC’s predecessor, the Armed Forces Institute of Pathology, was the place for pathologists and the greatest minds in the study of human disease to train. Navy pathologists have already used similar data to create a prostate cancer scoring algorithm that is better than a panel of 29 pathologists.

Some predictive applications have already been launched along these lines. The Military Readiness Assessment Tool (MRAT) provides monthly, rapidly-accessible, health readiness-related metric reports on Army units using a systematic format that does not require local data management. A team of Naval Medical Center, San Diego (NMCSD) and Google researchers validated a deep learning algorithm for improving Gleason scoring of prostate cancer. DHA epidemiologists and data scientists routinely leverage the Military Health System Data Repository, which represents clinical data from 9.4 million patients collected from 260 facilities and dozens of sources stretching back over a decade. The Applied Proteogenomics Organizational Learning and Outcomes (APOLLO), led by Dr. Craig Shiver of the Murtha Cancer Center, picks up where the NCI’s Cancer Genome Atlas (TCGA) left off, building a multi-source library of proteogenomic data from 8,000 cancers. Global Emerging Infections Surveillance (GEIS) global infectious surveillance: 140 sites around the world from Georgia on the Black Sea to Ghana, Thailand, Peru, and Guam, monitors primarily viral diseases with 140 partners around the world informing current operations.

Revolutionizing military, veteran, and dependent healthcare can turn into a major point of strength in the military’s relationship with critical players in the Silicon Valley tech community. The department has significant sources of data of intrinsic commercial value, some of which are addressed herein, and it is in the nation's best interest to capitalize the advantage these data resources represent. Big Data is what Silicon Valley does best, and the potential for spillover into civilian healthcare systems is vast.

Military medicine stands at the nexus of supreme technical challenges of life-or-death impact and tackles these challenges for warfighters. These challenges have broad global implications that technology companies are intensely interested in, which should not be overlooked by either side. These challenges include infectious diseases, search and rescue, medical evacuation, and survival in spartan environments, ranging from cold mountain warfare at altitude to desert warfare in open plains. Military medicine stands at the intersection of these national interests in a way that few segments of any industry can hope to.

The concerns that always come up in this context is the security and privacy of the data sets. Fortunately, protected health information (PHI) can be secured by policy at unclassified levels, and a growing collection of agencies and potential academic and industrial partners are credentialed and highly competent in both the necessary security procedures and active de-identification. Differential privacy and PHI redaction algorithms are already industry standards in creating robust training data sets.

None of these benefits can be realized without overcoming some challenges. The first major challenge is designing technical solutions that can be smoothly implemented and integrated into clinician practices and patient care. Behavioral change can have far greater impacts as a result of digital health, but changing habits is much easier said than done. This is also where the DoD has an advantage. Military culture makes behavioral changes easier to roll out, as the chain of command is a natural structure to implement and mandate changes in personnel, given those changes are well founded in science. This can apply to both the practitioner and the patient. This is not to say that some degree of user training, feedback, and further development is not needed, but military adopters tend to be easier to work with if a command directive is in place for them to lean into a new technology.

The other challenge comes from an advocacy and ownership problem. Every commander in the military would benefit from having a healthier, happier, and more available force (and healthier and happier military dependents), so effectively, the entire DoD is their constituency. The problem is there is no single entity has complete ownership over the resources (funding and data) needed to develop AI methods and worklows. Challenges in the Federal Acquisition Regulation’s execution timelines make methods such as other transition authority under 10 U.S.C. 2371b a desirable path to acquisition and prototyping, such methods allow much more agile means of acquiring technology and have been used by organizations such as Defense Innovation Unit with great success. As with any large organization (particularly in industries that pre-date the internet), there is also a misalignment in data owners, data stewards, and the beneficiaries of the analysis. High-level, sustained engagement and empowerment will need to happen for us to be successful. This will come with cost – in dollars, political capital, and organizational discomfort. But the costs of inaction are far greater; we in the Department of Defense have a greater obligation to give service members the best healthcare possible, and we have an obligation to the taxpayer to use their money in most fiscally responsible way possible. Unifying the largest clinical dataset in the world allow us to achieve both objectives.

Ryan Kappedal is a Lieutenant Colonel in the US Air Force and a Product Manager for Artificial Intelligence Systems at Defense Innovation Unit.

Niels Olson is a Commander in the US Navy and the Laboratory Medical Director at US Naval Hospital Guam.

Disclaimers for Niels Olson: "The views expressed in this online publication reflect the results of the authors(s) and do not necessarily reflect the official policy or position of the Department of the Navy, Department of Defense, nor the U.S. Government. I am a military service member or federal/contracted employee of the United States government. This work was prepared as part of my official duties. Title 17 U.S.C. 105 provides that ‘copyright protection under this title is not available for any work of the United States Government.’ Title 17 U.S.C. 101 defines a U.S. Government work as work prepared by a military service member or employee of the U.S. Government as part of that person's official duties."

Read more in the CNAS Military, Veterans and Society Program's "Supporting the Military Community" commentary series.

Image credit: Google Brain Team