Mount Sinai Hospital deployed an artificial intelligence model to look into a database of patient records and see what it could discover from the data. The result was an ability to predict patients who would be most likely to develop hypertension, diabetes, and schizophrenia.19 There are other initiatives in place across hospitals and departments worldwide. The ultimate question is how useful they will end up being—how reliable the results will be over time, and how trusted the systems will turn out to be by the doctors on the front lines.
Using Crowdsourcing to Track the Flu
The systems so far discussed in this chapter center around individual patient records and lab data being mined to find patterns that human beings simply can't detect, at least not with the same speed and reliability. Beyond the individual level, we can also look at broader populations in order to track infectious diseases like the flu. If we can harness environmental data—whether the people around you are sick or not—in an effective way, we can potentially detect your illness earlier (thus leading to better, cheaper, easier treatment) or help you avoid getting sick in the first place by keeping you from the epicenters of disease. We can stack the deck in our favor using population‐level information.
GlaxoSmithKline collaborated with MIT Connection Science in 2017 to launch Flumoji, a real‐time crowdsourced tracking engine that the website FiercePharma called “Waze for the flu.”20 The app attempted to use changes in users' activity and social interaction patterns—combined with flu‐tracking data from the Centers for Disease Control and Prevention (CDC)—to find outbreaks more quickly than traditional methods had been able to.
Flumoji was unique in using activity data as a proxy for wellness, but other crowdsourced engines have attempted to make similar predictions, including Flu Near You, which collaborated with Science Friday on NPR to track influenza outbreaks around the country.21 The problem with legacy tools like the CDC's “Flu View” report is that CDC data has a gap of almost a week between diagnosis and when physician reports hit the system and are able to be analyzed.22 Shortening that gap would save lives—allowing hospitals and doctors' offices to order sufficient supplies when a new wave of cases is expected, and giving people warnings to keep kids or the elderly home, or to make sure to get the flu shot.
For a number of years, Google tried to track the flu by using Google search activity data. But the company failed in 2015 to predict the flu season's peak.23 The Weather Channel used social media activity to create a map of flu activity—but this was designed to be illustrative, not predictive.24 The issue with all of these tools is whether they can prove reliable enough to move from merely interesting to actually useful in a way that changes behavior, treatment, and outcomes. We are likely not there yet.
A study in BMC Infectious Diseases looked at a range of crowdsourced flu‐tracking systems, including Flu Near You, and found that while these systems can add some information at a broad geographic level (national or regional), once you drill down close enough the correlation between systems decreased and the value was questionable.25 But it's at the neighborhood level where predictions can really make a difference in changing behavior and treatment…so it remains questionable whether these systems are truly adding value.
Stopping the Spread of Illness with Data Is Hard
I talked to Julian Jenkins, now at Incyte, but who spent seven years at GlaxoSmithKline and was working closely with MIT on the Flumoji project back when it launched in 2017.26 The most exciting piece of it, Jenkins explained, was trying to find those new biomarkers that mattered, the new layers of the layer cake that would make a difference. With the app running, you could track whether bedtime changed the night before illness set in, if the pattern of what someone was looking at on Facebook or Twitter changed, or if something about the person's online presence signaled an illness even before there was a search for whether CVS was open. Could we learn someone had the flu even before they realized it themselves?
Jenkins believes there is much data out there that we're not yet effectively mining. Could our TV set, for instance, tell us something we don't know about ourselves, from our viewing patterns, from how much we're moving around while we're watching? Could the GPS on our phones indicate the kinds of restaurants we're visiting, or the kinds of stores we're shopping in—and could that add useful information to an understanding of our lives from a medical perspective?
This kind of thinking isn't limited to the flu. India Hook‐Barnard, director of Research Strategy and associate director of Precision Medicine at the University of California, San Francisco, told MobiHealthNews that, on a population level, we can use activity data to better understand all kinds of conditions.27 Are people in a certain neighborhood at greater risk of cancer, diabetes, or another illness? Is it because of their access to healthy foods? Is it because of the availability of medical care? Is it about the availability of effective messaging about health? About environmental factors we may not have previously considered? “Knowing that a given population is at a greater risk,” Hook‐Barnard told MobiHealthNews, “you can do earlier diagnostics and screenings for certain diseases and then be able to be more effective in your surveillance of those people.”28
This keeps people healthier and allows for much earlier intervention. Yet, at the same time, Julian Jenkins believes that while technology can help us understand and predict disease in so many ways, there are many significant barriers still in place, not the least of which is broad participation right now in efforts like Flumoji and others. To get people to even download the Flumoji app was a challenge—and that challenge becomes even bigger when we're talking about data centered around clinical trials, where a very small percentage of patients will ever be involved. The lack of interoperability—for the Flumoji app, needing to develop a front‐end app for multiple platforms (Apple, Android, etc.)—as well as the lack of data sources being able to communicate behind the scenes (a problem we'll talk much more about later in the book)—is a tremendous challenge we still need to overcome.
When it comes to the hospital‐based projects we talked about earlier in this chapter—like Sepsis Watch and ColonFlag—a big part of the work is ensuring that the health records data is fully usable and can be reliably accessed in an efficient, streamlined way. It's a huge problem to try to integrate new technology with legacy data systems, and it's not a problem that's going to disappear anytime soon. The mere ability to have data entered automatically rather than requiring someone to manually do it—a process that takes significant amounts of time—is still something we are working on in a lot of situations, so real‐time analysis isn't always possible, even if the algorithms are there.
When it comes to patient equations and precision medicine, if we're being realistic, the flu isn't the holy grail. It would be terrific to ease people's suffering—and, in the case of sepsis, it would be tremendous to save people's lives through better prediction—but where precision medicine is already making the most difference (and where hopes are highest) is in areas like cancer, where the stakes are so high, and the current landscape is so far from ideal. Death rates from cancer have declined by 27% since 1991,29 but much of that progress has been from lifestyle changes like fewer people smoking. There is hope that precision medicine can lead to more than just incremental change, but rather to revolution in detection and treatment. It is a more complex problem than any we've looked at in this section of the book, but patient equations offer hope here that hasn't been seen in decades.
Notes
1. Yang Li, “What Should We Learn? Hospitals Fight Sepsis with AI,” IEEE Signal Processing Society, November 5, 2018, https://signalprocessingsociety.org/newsletter/2018/11/what-should-we-learn-hospitals-fight-sepsis-ai.
2. Cara O'Brien, MD, and Mark Sendak, MD, “Implementation and Evaluations of Sepsis Watch – ICH GCP – Clinical Trials Registry,” Good Clinical Practice Network, 2019, https://ichgcp.net/clinical-trials-registry/NCT03655626.
3. Laura Ertel, “Buying Time to Save Sepsis Patients,”
Duke University School of Medicine,” June 4, 2019, https://medschool.duke.edu/about-us/news-and-communications/med-school-blog/buying-time-save-sepsis-patients.
4. Cara O'Brien, MD, and Mark Sendak, MD, “Implementation and Evaluations of Sepsis Watch – ICH GCP – Clinical Trials Registry.”
5. Yang Li, “What Should We Learn? Hospitals Fight Sepsis with AI.”
6. Mark Sendak et al., “Leveraging Deep Learning and Rapid Response Team Nurses to Improve Sepsis Management,” 2018, https://static1.squarespace.com/static/59d5ac1780bd5ef9c396eda6/t/5b737a1903ce645e7ad3d9a2/1534294563869/Sendak_M.pdf.
7. Ibid.
8. Laura Ertel, “Buying Time to Save Sepsis Patients.”
9. Will Grant, “Eric Poon's Boundary‐Pushing Use of Technology at Duke Health,” American Healthcare Leader, February 4, 2019, https://americanhealthcareleader.com/2019/poon-tech-patient-care/.
10. Eliza Strickland, “Hospitals Roll Out AI Systems to Keep Patients From Dying of Sepsis,” IEEE Spectrum: Technology, Engineering, and Science News, October 19, 2018, https://spectrum.ieee.org/biomedical/diagnostics/hospitals-roll-out-ai-systems-to-keep-patients-from-dying-of-sepsis.
11. Mark P. Sendak et al., “Sepsis Watch: A Real‐World Integration of Deep Learning into Routine Clinical Care,” Journal of Medical Internet Research, June 26, 2019, https://www.jmir.org/preprint/15182.
12. M. C. Elish, “The Stakes of Uncertainty: Developing and Integrating Machine Learning in Clinical Care,” 2018 EPIC Proceedings, October 11, 2018, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3324571.
13. Ibid.
14. Cara O'Brien, MD, and Mark Sendak, MD, “Implementation and Evaluations of Sepsis Watch – ICH GCP – Clinical Trials Registry.”
15. Will Grant, “Eric Poon's Boundary‐Pushing Use of Technology at Duke Health.”
16. “3 Considerations for Adopting AI Solutions,” American Hospital Association, 2019, https://www.aha.org/aha-center-health-innovation-market-scan/2019-01-08-3-considerations-adopting-ai-solutions.
17. Bill Siwicki, “Hospital Cuts Costly Falls by 39% Due to Predictive Analytics,” Healthcare IT News, April 12, 2017, https://www.healthcareitnews.com/news/hospital-cuts-costly-falls-39-due-predictive-analytics.
18. Paul Cerrato and John Halamka, “Replacing Old‐School Algorithms with New‐School AI in Medicine,” Healthcare Analytic News, April 5, 2019, https://www.idigitalhealth.com/news/replacing-oldschool-algorithms-with-newschool-ai-in-medicine.
19. Thomas Davis, “Artificial Intelligence: The Future Is Now,” ProCRNA, April 21, 2019, https://www.procrna.com/artificial-intelligence-the-future-is-now/.
20. Beth Snyder Bulik, “GSK and MIT Flumoji App Tracks Influenza Outbreaks with Crowdsourcing,” FiercePharma, January 28, 2017, https://www.fiercepharma.com/marketing/gsk-and-mit-flumoji-app-tracks-influenza-outbreaks-crowdsourcing.
21. “Tracking The Flu, In Sickness And In Health,” Science Friday, 2018, https://www.sciencefriday.com/segments/tracking-the-flu-in-sickness-and-in-health/.
22. Laura Bliss, “The Imperfect Science of Mapping the Flu,” CityLab, January 30, 2018, https://www.citylab.com/design/2018/01/the-imperfect-science-of-mapping-the-flu/551387/.
23. Ibid.
24. Ibid.
25. Kristin Baltrusaitis et al., “Comparison of Crowd‐Sourced, Electronic Health Records Based, and Traditional Health‐Care Based Influenza‐Tracking Systems at Multiple Spatial Resolutions in the United States of America,” BMC Infectious Diseases 18, no. 1 (August 15, 2018), https://doi.org/10.1186/s12879-018-3322-3.
26. Julian Jenkins, interview for The Patient Equation, interview by Glen de Vries and Jeremy Blachman, March 24, 2017.
27. Bill Siwicki, “What Precision Medicine and Netflix Have in Common,” MobiHealthNews, May 22, 2017, http://www.mobihealthnews.com/content/what-precision-medicine-and-netflix-have-common.
28. Ibid.
29. Stacy Simon, “Facts & Figures 2019: US Cancer Death Rate Has Dropped 27% in 25 Years,” American Cancer Society, January 8, 2019, https://www.cancer.org/latest-news/facts-and-figures-2019.html.
7
Cancer and Phage Therapy—Crafting Custom Treatments Just for You
In 1971, U.S. President Richard Nixon announced a war on cancer, and yet almost 50 years later, there has been relatively little progress made compared to expectations. The biggest challenge is the complexity of the disease—or, as we ought to say, diseases. Cancer isn't one disease, it's many—and, in fact, in some ways, it's a different disease for just about every patient, with a number of individual factors contributing to every tumor, and every case. It is often hard to find evidence of cancer before it is too late to cure, treatments that work for one person's illness often don't work for another's, and the cancer itself can change over time to become resistant to a treatment that's working and require an entirely new strategy. When we talk about personalized medicine, it doesn't get much more personalized than cancer treatment—and when we talk about the need for precision, the stakes are rarely as high as they are here.
Later in this chapter, we'll talk about another very personalized approach to a complex, individual problem: phage therapy, or using custom bacteriophages to treat serious bacterial infections. But we'll start our exploration of personalized patient equations with cancer. On the diagnosis end, there are certainly intriguing developments—just as one example, a startup, Cyrcadia Health, developed a patch for women to wear under their bra in order to track breast tissue temperature, alerting wearers to see their doctor for further examination if there is a change in pattern.1 As another example, researchers are looking at detecting cancer via breath test, under the theory that cancers change the pattern of molecules in the breaths we exhale—and that in fact there may be molecular fingerprints within the data that can detect not just cancer in general but particular cancers, more quickly than traditional tests.2
But perhaps the most interesting breakthroughs concern treating cancers that have proven resistant to traditional treatments, and where the combination of genetic information with proteomics—the proteins in a patient's tumor sample—have opened the door to new approaches. Former U.S. Vice President Joe Biden is quoted by Dr. Jerry Lee of the University of Southern California (who you'll hear a lot more from in Chapter 9) in a piece about the global Cancer Moonshot effort as saying, “it's like the genes are the full roster of a basketball team but the winning strategy comes from finding out who their starting lineup is. The proteins are the starters you're going to play against—the five you are going to have to defend against.”3
With information about those proteins combined with genomic information, we've been able to move from talking about breast cancer, for example, not as one disease, but one with subcategories, like triple‐negative (meaning the tumor is absent the three most common receptors for cancer growth: estrogen, progesterone, and the HER2/neu gene), allowing us to customize treatments and get the kinds of results we've been dreaming about since Nixon's declaration of war.
Changing the Way We Look at Cancer
In the old way of looking at things, cancer is simply a disease of uncontrolled growth and uncontrolled cell division. But over time, we've learned that it's more than that. A successful cancer needs to build its own infrastructure. It needs to stimulate the growth of new blood vessels. It needs to invade. If we can cut off its ability to build what it needs to keep growing, we can stop it in its tracks. I like to think of this as a cybersecurity problem. What we need to solve it is an air‐gapped computer—meaning it's not able to connect to the outside, no attached network peripherals, no USB drive, no access to the Internet. We need the cancer to become an isolated system.
Even in the computer realm, this is harder than it seems. Data can enter and leave even through a power cord. And if you're connected to a wireless network, then you're pretty much dead. Cancer has even more vulnerabilities it can exploit, lots of ways it can get what it needs and ultimately destroy its host. Traditional therapies have been like
carpet‐bombing, with all of the associated collateral damage that chemotherapy can cause. But if we can target angiogenesis, for example, we can limit the cancer's ability to grow. If we can successfully stop its progression without destroying healthy cells in the process, it's a huge victory.
Newsweek announced in 2008 that “We fought cancer…and cancer won.”4 Indeed, there is so much left to figure out. With undiagnosed disease, data can get us only part of the way there. With genetic sequencing we are able to diagnose about 40% of patients, and for only about 40% of those are we able to find a treatment that might work. It is sobering to realize that we can't (yet) cure cancer with math. Cancers are excellent at inventing ways to get around whatever particular pathway we are trying to block. Even with targeted therapies, we eventually lose. So we need not just a targeted first‐line therapy but a second‐line and third‐line therapy as well. We need to be ready before the cancer forces a change. We need to track the cancer closely enough to beat it—and transform the disease from something deadly to perhaps something chronic, where we won't necessarily ever utter the word “cure” but where we can keep finding new ways to press the pause button.
In that spirit, the more we can measure, and the more intelligence we have about not just the right treatment but about whether or not the current treatment is still working, the better shape we are in. Judging whether a treatment is still effective is not always easy, and certainly not on a real‐time basis. Traditionally, we need to cut and poke, or at least perform a time‐intensive scan—but now, more and more, sensors can give us that continuous multivariate view we need, if not to measure exactly what we're looking for, then at least to measure potential proxies for it.
The Patient Equation Page 11