What's interesting is that PatientsLikeMe launched in part because of Jamie's sense that the great leaps forward in treatment—for lots of conditions—weren't going to happen without examining the kinds of individual, personal experiences that were being ignored by traditional clinical trials because they lacked the objective and mathematical rigor that trials required. The thought was that by harnessing enough of that information, with big enough sample sizes, the scientific flaws would work themselves out. And while that reasoning made sense in 2004, when PatientsLikeMe launched, we're now in the position to have the best of both worlds. We can still analyze these rich personal accounts—these individual experiences and feelings—but sensors have made it possible to bring enough rigor to the measurements that we can in fact incorporate them into clinical trials without compromising on the scientific standards.
Heywood's frustration is that the layers we've always looked at—and assumed to be sufficient for developing drugs and treating disease—just aren't enough. If we want to optimize health—and not just spot‐treat specific illnesses—he believes we need to look much more at things like our circadian rhythm, parasites, inflammation, and the biome of flora and fauna in our environment, just as examples. His list goes on from there: our social relationships, the virome, the anxiety and mental health issues that come with chronic disease. And if we could gather all of these layers of information from everyone on the planet—which, more and more, we can—and track them against health outcomes, we would see connections that we've missed for thousands of years.
There are obviously implementation issues, even if you accept that this is a worthy direction to explore, and I'll talk more later in the book about harmonizing data sets and integrating real‐world experiences into the clinical trial framework. (“We can't make health data meaningful until it's digital, and right now it's not digital,” Jamie says.) At its core, PatientsLikeMe is seeking to find those same inputs for patient equations that we're looking for, but instead of coming at it from the point of view that many of us in life sciences have—“what can we do inside of clinical trials?”—they're looking at what we can do outside of a research environment.
Talking to Jamie hammers home what is one of the most important points about patient equations: useful models for diseases are multivariate. There are so many influences, so many interacting elements that we have never considered before in the treatment calculation, not necessarily because we didn't think they might be important—though possibly that—but also because we couldn't measure them. Now we can, and we can measure them with even more rigor than Jamie Heywood imagined we could when he started PatientsLikeMe hoping to aggregate people's lived experiences in order to help patients make more informed decisions about how to approach their conditions, and how to optimize health and wellness.
Where PatientsLikeMe has had a great deal of success is in helping patients manage disease in pockets that research hasn't touched and where doctor intuition isn't necessarily accurate or robust. For instance, compared with a study interviewing 20 neurologists, the PatientsLikeMe recommendations for how to manage excess saliva for ALS patients were better. Comparing different ventilators, comparing the effects of different mental health treatments, or figuring out whether a multiple sclerosis patient can maximize quality of life by starting on medication, by changing diet, or by installing an air conditioner to better manage heat in the summer are all things that have always been theoretically measurable, but impossible to actually, rigorously measure. PatientsLikeMe set out to get around the measurement problem by collecting anecdotes. With the right technology, patient equations can do even better, and translate those anecdotes into real data.
“So many things are so much better than drugs,” Heywood says. “Things we are not even running studies on.” And the studies we are running, he says, are too narrowly focused on one dimension of illness, on the pathology of the disease state rather than on more broadly‐defined wellness. He is skeptical right now of the sensors and wearables we have—and, the truth is, despite my chest patch and my wristband, so am I. We're measuring a set of isolated variables right now, but the algorithms in the background are still at the level of needing more information to even know what we ought to be measuring, let alone how to act on those measurements, and how to combine them in useful ways and get real information out of them. “There's lots of signal, but not a lot of meaning yet,” Jamie says. Indeed, we still have a long way to go. But there are some big conceptual ideas in the rest of this chapter that we can think about as we look ahead, as we start to imagine a world driven by more accurate, more comprehensive, more predictive patient equations.
Incidentally, that potentially meaningful territory measurement I've talked about emerged from a patient anecdote—not on PatientsLikeMe, but from a conversation at Medidata. Barbara Elashoff (a former FDA statistical reviewer, and founder of a startup whose company was acquired along the way by Medidata) was a weekend‐warrior athlete.6 Her step counts would have been through the roof—until she broke her leg, ironically while jumping over a trench in a tough mudder‐style event. We were discussing how, if you had been measuring only her step counts, you would have seen them go to almost zero, almost instantly. However, alone, that wouldn't have told us much about her health. It could have meant she was hospitalized for some reason not related to a musculoskeletal injury, or just that she started a new job or a project that forced her to spend her days and weekends behind a desk.
But combined with other information—lifestyle questions, her blood chemistry, and cardiac markers—we were thinking about how we could use proxy data to determine that it was, in fact, an injury from exercise. And about how we could use that same data to plot her recovery. Again, one measurement alone won't do it, but a stream of data, over time, could tell us something important about her life. During that conversation, the idea of patient territory was born, when someone pointed out that steps may not be the right measurement of activity (since taking the bus should count for something). Apps were built, and algorithms for calculating and summing area from vectors were designed that could run in the privacy of a patient's own phone (with Barbara's husband and her company's co‐founder Mike furiously reading academic papers on monitoring wildlife and writing code to turn techniques from ethology into useful tools for human medicine). The future clinical utility of patient territory is yet to be proven—but the pursuit of it and ideas like it is continuing.
Changing the Frequency
Traditional medicine—and the traditional medical biomarkers that go along with it—typically involves discrete measurements taken at relatively distant timepoints. A physician sees a patient, measures certain things, prescribes a course of therapy, and then days, weeks, months, or maybe even years later takes the measurements again to see what the effect was on the patient. But the fact that these measurements are taken at single moments in time gives us very little understanding of how they might be changing, particularly when timepoints are so far apart that we may miss entire cycles of the value being elevated, continuing to go up, then coming back down, and eventually returning back to normal, just as an example. The measurement itself may be perfectly precise, but it's actually a very imprecise way to measure the range of values the patient lives with, or the rate of change of those values.
This isn't meant to condemn the physician. It is based on the very real limitation that we typically have only been able to measure things reliably about patients when they were physically present with a health care professional. We measure blood pressure and heart rate, or take blood samples, at selected moments in time—usually in a clinic setting. We potentially take action based on a perceived trend in those numbers, but there's inevitably an element of guesswork. Are we catching the patient only at moments when their heart happens to be in sinus rhythm, when in fact they're having significant episodes of atrial fibrillation when they aren't hooked up to an ECG machine? Is the patient experiencing “white coat hypertension,” with higher blood p
ressure readings in their doctor's office than they would have at home? Are we only flossing right before we go to the dentist, and is that changing his or her read on the condition of our dental hygiene? Is whatever measurement we're looking at a straight line between two timepoints, or a punctuated equilibrium, with periods of stability and then sudden spikes that we might miss?
Sensors allow us to move from the staccato rhythm of traditional medical care to a continuous stream, from a low‐frequency data environment to a high‐frequency one, and one where real‐time access to that data allows us to act instantly. Ideally, we can now see right when alarming trends begin rather than hoping we'll catch them before it's too late. Plus, with a well‐instrumented patient, we can be measuring the things we don't yet know we need to measure in order to find evidence that we're not yet even looking for. Taking the atrial fibrillation situation as an example, normally it wouldn't be diagnosed until a stroke or some other potentially‐damaging event, because we aren't typically on heart monitors without a previous episode. But if we can catch those patients most susceptible and start them on blood thinners before that first stroke, we can save lives. (As one example, a 13‐year‐old boy's Apple Watch detected a heartbeat of 150 bpm, the boy was rushed to the hospital, doctors found that he was in atrial fibrillation, and were able to intervene before there was a problem.7)
When it comes to white coat hypertension, a study reported in the New England Journal of Medicine in 2018 found that 24‐hour blood pressure monitoring was more predictive of mortality than if it was merely measured in the clinic.8 This begs the question of what might change, with regard to how we treat hypertension, and how we measure how effective those prescriptions are, if we could measure all of our blood pressures 24 hours a day.
These high‐frequency data streams also allow us to break the need to have the patient physically present with a health care professional to reliably measure things about them. The doctor is just as much a beneficiary as the patient here, as we try to understand the nature of diseases. Patients can be monitored continuously, not just at those discrete moments every six months when they have an appointment, with data alerting physicians if something requires their attention, a change in a critical measurement, or a progressing trend that could be missed. Of course you can't do a full‐body PET scan every day. But a passive device working in the background can collect data 24/7. And on a patient's end, high‐frequency feedback can help optimize medication dosing and manage conditions as effectively as possible, alerting them when something is just beginning to go off course, rather than when damage has already occurred.
Patient equations in this high‐frequency world move us from the stock pages in the business section of the next morning's newspaper to the constant scroll on the Bloomberg terminal. The newspaper is for amateurs. The Bloomberg terminal is for professionals. We don't have to wait until the next day. We can see the patient's condition in real time, the terms of the equation changing as the patient moves through the world.
Reverse‐engineering the Critical Layers
Right now, the equations we use in diagnosis and treatment are often univariate—one variable, simple if‐then formulas: if cholesterol is greater than 180, then prescribe a statin. This is not because we truly believe 180 is a magical cutoff between safety and danger. It's not even because we believe cholesterol is the perfect solo measure of whether someone needs a statin. It's because it's the best we can easily do. It's a useful abstraction of a problem, allowing us to treat people at a societal scale. However, that's the kind of model that I'm hoping this book can help move us beyond. Technology now gives us the ability to add new terms to the equation not just because they're easy to measure but because adding them makes the equation work better, giving us the ability to diagnose and treat more effectively.
The way we'll get there is by instrumenting patients, storing those digital biospecimens, and replaying the past. (And it's cheap—we can store digital streams at a far lower cost than keeping tumor specimens on ice.) It may be, as Jamie Heywood might argue, that the micronutrients in our diet have serious implications that we simply don't realize right now. We need certain metals at trace amounts in our bodies, but we're not monitoring that intake. Would it make a difference for some of us if we did? At the same time, what about those of us eating too much? Is, say, mercury toxicity affecting our health in subtle ways we're not picking up on? We can measure these things and then, in a few years, look at outcomes across a population. We may uncover a real connection. We may not, but then we'll have that added piece of knowledge. Do this across a range of variables—sleep, resting heart rate, performance on brain‐training games, anything else we can imagine and track—and we will start to be able to reverse‐engineer the best‐fit, most accurate patient equations out there.
Too often, discoveries are therapeutic accidents. Former Senator John McCain's deadly glioblastoma was found incidentally during surgery on a blood clot. It's not an atypical story. Therapeutic accidents aren't a good strategy for increasing health in a population. They're a lesson that there's something we should be finding in a different way. We tell patients that a drug works for, say, 80% of the people to whom it's prescribed. But how do we know who is going to be in that 20% of nonresponders, in a perfect world even before we write the prescription? Right now, we probably don't. That doesn't mean the data isn't out there somewhere, waiting to be discovered.
There is, undoubtedly, a point of decreasing returns. There are infinite things we can measure, but perhaps only a few that will make significant differences in most cases. We're not there yet. We don't yet know the five terms or the 500 terms that will make the biggest difference in most of these equations. We'll get there, and then we can start simplifying—then we can say, “well, if we're worried about long‐term outcomes, we just need to look at hemoglobin A1c, total cholesterol, blood pressure, and a self‐assessment of diet and exercise, because even a thousand other variables won't add more than 1% accuracy to our overall score.” We know that if you manage your cholesterol and blood pressure, then there's a much lower chance of having a stroke or heart attack. But are there things we can supplement those tests with to get us to an even better prediction? How much benefit is there to adding the resting heart rate, the calcium score, or the continuous ECG feed? Are there other multidimensional ways to look at cardiac health that will have greater mathematical impact on our long‐term livelihood?
We know that if we keep anyone on a Holter monitor for long enough, we will find abnormal moments, unusual blips. But how do those compare to what we see on a problematic ECG? What is the spectrum of good to bad readings, and how much data do we need to plot someone's results accurately on that chart? Or does it depend on other variables? Do two readings that look the same mean very different things depending on other measurements that we perhaps haven't yet identified as being related? The better we can understand how to draw lines between readings, the better predictions we'll be able to make about disease, the better decisions we can make about optimal care.
We certainly won't be able to predict a heart attack based on step count alone, but maybe we can get closer. Maybe we can't. Some of the cleverness of the new technologies (some of which we'll examine in Section 2) is in figuring out what not to measure, and separating which measures are useful and which aren't. We are just beginning the journey from univariate to multivariate approaches to modeling disease. Right now, to most of these questions we're asking, the answer is that we just don't know. One day, we will.
The Cognitive Dimension
I think it's almost impossible to overstate how much we are missing from our disease models right now when it comes to cognitive measurements. We may think of the body and the brain as separate spheres, that your state of health and your state of mind are independent actors. But there is so much underappreciated connection between the two—optimism and pessimism related to hormone release, changes in metabolic rates, and so much more.
The connections b
etween brain and behavior are obvious and easily observed. We decide to walk from point A to point B. We consciously or subconsciously fidget. Our sympathetic and parasympathetic nervous systems make us run (or hide), yawn, and sleep. Without getting into any metaphysical discussion of the mind or the soul, it's the inarguable truth that inputs from our sensory organs stimulate the voluntary and involuntary movements of our muscles, and our autonomous physiological functions work with the same types of feedback loops.
There is mounting evidence of how large a role behavior and cognition play—in ways I find unsurprising—in the progression of disease. A landmark study in 2018 suggests that prescribing a Fitbit can affect cancer survival.9 Is this the result of patients moving more, and thus keeping their metabolic rates up? Maybe. There are well‐known connections between the immune system and our metabolic rates, the reasoning at the very least circling around the idea that there is energy required to produce an immune response, and so our ability to supply that energy has an impact on the immune system. There are also links between our immune system and cancer, so it is not a stretch to hypothesize that more active patients might have better outcomes.
The Patient Equation Page 5