3. Clinical Application

Ultimately, the goal of the individualized approach is to improve health and wellness for each person who applies it.  However, as the saying goes, 'the secret to happiness is managing expectations', and so it is important to understand what scenarios this approach holds the most promise, and where it is less likely to provide additional guidance beyond what has come from standard approaches.  In this post, we'll start by walking through an illustrative case of how the process would occur in a clinical setting, then discuss some details and challenges for actual implementation.  

Clinical Scenario: A patient with recurrent migraines

Becky has had migraine headaches for as long as she can remember.  Although the specific timing and patterns have changed over the years, she has generally had a headache every couple of weeks of various severity, with some causing her to need to leave work and just sit alone in a dark room for up to a day.  She has tried various medications over the years, with some working for a time before they stop working, and others not working at all or seeming to make things worse.  She has also had some of her own theories about lifestyle factors that seem to bring them on, such as heavy drinking or dehydration.  On the whole, however, she has continued to have them and feels at times as though her life revolves around a problem she cannot totally understand or control. 

Now, how would this situation be managed currently in our medical system?  Let's assume for a minute that Becky has insurance, and can afford to see her primary care provider to discuss rather than having to go to the Emergency department.   The first step is obviously to make a diagnosis, and make sure that the headaches are not something requiring urgent care, such as a tumor or intracranial bleeding.  After questioning, the provider comes to the conclusion that these are likely recurrent migraine headaches.  That's helpful, and obviously good to rule out potentially life-threatening things, but it doesn't automatically make the headaches go away.  This is where a lot of modern medicine tends to break down, because the next step is for our provider to apply a treatment that has been studied in many people other than Becky.  Further, if the provider is using 'evidence-based medicine', then he or she is actually not basing this treatment individual response or on his or her own personal experience treating migraines, but on studies averaged across thousands of individuals, many of whom will have a widely different response. 

The provider asks Becky if she can think of any clear triggers, advises her to avoid them, and then prescribes the medication and recommends that she follow-up in 3-6 months to see how it's going.  There's no monitoring plan or system, and no method for quantifying the severity or frequency of the headaches.  An astute patient or provider might use a log of the headaches, but there's no analysis on the back end, more of a qualitative look during the brief return visit.  On Becky's return to the clinic, based on her subjective opinion she tells the provider whether the treatment is 'working' or not.  If not, then the provider can recommend another treatment for trial and error, with Becky returning later without any data collected or analyzed to say whether things are 'going well or not'. 

Of course, this is not an ideal situation.  And yet, for many recurrent, incurable, medical conditions, this treatment approach is the best modern medicine has to offer.  Many would offer that because over the past 30 years we've started analyzing data across thousands of people rather than using a single doctor's anecdotal experience, we've made a major breakthrough from the 'dark ages'.   Well, we can do better.  In an age where people can be reached 24/7 because they carry around a phone, where heart rate monitors collect real-time data on every heart beat every minute of the day, and when a smart phone processor has more computing power than the computers that put a man on the moon, we can do better.  In this post, we'll describe the individualized medicine approach that we believe to be the future of medicine.

The individualized approach to our patient   

We start the individualized process with Becky by collecting data.  We have her wear an activity and sleep monitor, a light-sensing skin patch to track sun exposure, and a heart rate, breathing rate, and body temperature monitor implanted under her skin.  She uses a smart water bottle to track her fluid intake, a smart toilet that tracks her body weight and urine output, and a home monitor to track home temperature, light exposure, and noise.  She also enrolls in a passive data collection program with her credit card and grocery store loyalty cards, which allow purchase data to be filtered and uploaded to her data analysis platform.  Finally, she uses a simple user interface to keep track of her headaches (timing and severity) on her smartphone.  Data is collected over a year-long period, during which time she has approximately 18 migraine headaches of various severity.  

An analysis is run using a Bayesian modification to a Hidden Markov model that incorporates priors from other individuals of her age and gender with migraine headaches, and several lifestyle factors are found to correlate with her headaches at differing lags of time.  Notably, the model implicates the amount of sleep she gets the week before migraines, which it also notes is directly correlated with how early she goes to sleep.  In other words, based on the data analysis, it appears that when she goes to bed too late too often in a week, she is at increased likelihood of having a migraine headache the following week.  She sits down with her individualized medicine provider, who discusses the results, and they agree on a strategy to try to test this hypothesis about sleep by going to bed at a specific time each night.  

She goes home and tries to keep to the early bedtime program, which she is able to for the most part, although occasionally she has no choice but to stay up for social gatherings, etc.  Data collection continues, and she returns 3 months later, where her intervention is examined.  She has had migraines, although based on seasonal trend adjustments, she has statistically decreased the number of migraine headaches she is having with this intervention.  The model is re-run, and it verifies that sleep is still a major factor in the migraines.  She discusses the situation with her provider, stating that she has done the best she can with the sleep based on her schedule.  Her provider acknowledges this fact, and turns on a feature in her user interface that will alert her when her risk of a migraine is higher than normal so that she can take a medication to try to prevent one from occurring.  She leaves and over the next 6 months has only one migraine (down from about 10-15 a year).  

Obviously, this approach, certainly some of the sensors and monitors described, is not currently available for patient care.  However, in many ways the process of collecting data, discussing it with an expert provider, selecting a treatment plan, and then measuring the response, has been occurring in healthcare since the earliest days.  The key difference here is that rather than using trial and error and gestalt, we are measuring hard data and making and testing statistical hypotheses.  In the figure above, this approach is generalized, including the option to visualize data and make qualitative inferences in addition to quantitative.  There is great power in numbers and the objectivity that comes with hard data.  However, more work is needed to make this clinical model a reality.   Below we discuss some of the clinical challenges involved. 

Types of Outcomes

As was discussed in detail in a previous post, the main factor in determining the type of outcomes we can use the individualized data analysis approach for is statistical power.  The outcome we are interested in needs to occur enough times for us to make a statistical inference, otherwise there will be too much uncertainty around our estimates.  An outcome that only happens once or twice in a person's lifetime, such as a heart attack or stroke, clearly does not provide enough information for us to build an individualized data model to predict risk factors and prevent it from happening again in the future.  On the other hand, a recurrent outcome, such as a migraine headache, panic attack, or back spasm, if it occurs frequently enough, can provide the amount of information for us to make and test statistical hypotheses about what other factors could be involved.  Even better, an outcome that can be measured every day, such as weight or amount of sleep at night, can go even further towards drawing statistical inferences since we would clearly have enough information to make a hypothesis.  

Medical vs. Wellness

Although the lines are somewhat blurred, there is a key differentiation to be made between a medical application and a wellness application.  Medical applications, like medical devices and drugs are under the jurisdiction of the Food and Drug Administration, and require careful testing prior to public release.  Wellness applications do not.  This distinction is important, since my life would not necessarily be in danger if my Fitbit does not track my steps accurately, but it would if my home glucose monitor failed to provide accurate measurements of my blood sugar.  A device or app aimed to inform healthcare decisions needs to be evaluated by the FDA (usually with clinical trials) to determine safety and efficacy.  

This distinction seems to make sense, certainly on the application level; however, data is data, and what would it be if I used data from a non-FDA, wellness-based device (i.e., a Fitbit sensor) to draw some inference about a medical condition?  Am I only allowed to use FDA-approved sensors, even if the data that they provide could be useful in an algorithm?  Does the outcome matter?  It seems pretty benign to use the data from a non-FDA approved activity monitor to test association with the amount of sleep each night, but what happens if I run the model and decide to take a sleeping pill based on a prediction of less sleep?  These are not easy questions to answer, but they will be important to keep in mind as we move from a healthcare model based only on treating present disease toward one that seeks to prevent disease.  

The Role of Providers

Finally, as was evident in the example above, for the individualized process to work requires that the provider, in addition to the patient, buy into the inferences generated from the model.  It is important to recognize that this is a major change with how physicians have generally been trained to interact with patients.  We do not run statistical models based on the information patients provide; in fact, most practicing physicians have minimal understanding of statistics beyond what was drilled into them in medical school for interpreting the results of clinical trials.  Myself and others have argued that there is real value in application of Bayesian or hidden Markov models in healthcare, since both approaches are quantifying concepts that we apply outside a formal mathematic framework.  Pretest probability is the term heard often on the clinical floors about whether or not to order a test (ordering a test when the pre-test probability is low is not a good idea since it only creates doubt, not confirmation).  An individual occupying various states of risk, like the states of a HMM, is a key concept taught early in medical school about the reason we need to 'lay eyes on the patient'.  'Sick or not sick', is the intuition taught, which could be easily assigned to a two state HMM, which can be further modified during the treatment course.  It is not unrealistic to believe that providers can take the additional steps of interpreting an HMM model or Bayesian analysis, especially if they do not need to completely understand the math behind them.  Nonetheless, engaging providers in the individualized process is a key component for success, and one that cannot be overlooked in implementation.  

Clinical validation

The final point we will address in this section concerns validation.   As we discussed above, in many cases, the development of evidence-based medicine was a major breakthrough in modern medicine as it created an objective standard by which new and existing treatments could be compared.  There are numerous examples throughout recent history in which a treatment that 'made sense' based on physiology or pathophysiology of disease, or that 'appeared to have worked before' in a handful of anecdotal patients, was found to either be ineffective or even cause harm when studied in a randomized controlled trial.  Today the barrier to implementing a new medication or procedure is often study in a large group of people in a randomized controlled trial; however, clinicians and researchers are realizing that this is not always feasible.  Alternative approaches, particularly using well-designed observational studies or cohorts are being explored, although these have significant limitations with bias and confounding.  

Of course, clinical validation of a single drug or procedure is one thing, but how do we validate an approach, with many different treatment pathways, patient characteristics, and outcome measures.  Investigators have attempted these studies, called outcomes or quality-improvement research, and have had some success identifying trends in treatment, although to date few of these studies have provided much in the way of new guidelines for practicing physicians.  The reason is one of complexity.  How do you account for all of these separate variables in a single study?  The curse of dimensionality tells us that we can't, and that adding more variables to a model only decreases the power to find anything at all.  The number of factors is too large relative to the number of patients, and often adding patients only adds additional factors that differ from the initial population. [If I expand a study to run in Africa as well as the United States, I need to account for the genetic and environmental differences in those populations as well, otherwise I will have a high likelihood of confounding]

So where does this leave us?  Can you do a study on the 'practice of medicine' at a high level?  Can we honestly randomize our patients into two groups, one of which we apply 'individualized medicine' and the other standard approaches, without bias?  The answer is that we presently have no answer.  There are some investigative methods that aim to split the difference, using approaches such as stepped wedge randomization in the emerging field of implementation science.  These approaches may have promise, although closer inspection raises a number of methodological limitations that would be unlikely to pass the rigor of serious epidemiological review.  At the end of the day, they may make us feel 'better' about using a given approach, especially if our patients seem happier and there do not appear to be extra complications or risks.  On the other hand, the placebo effect can be quite strong, and if we really care to know whether an approach works, we need better.  Hopefully as the Individualized Data Analysis Organization moves our own projects forward we will be able to fill in some of these gaps.  

Next, Part IV. Statistical considerations