1. Data Collection
I recently attended a big data scientific session in which the keynote speaker, a high-ranking software engineer and data scientist from a major Silicon Valley company pointed out that it was embarrassing how little biological information we in the medical field use when making patient decisions. He had a point, which is that when you consider the complexity of the human body, with countless circulating metabolites, proteins, RNA, and DNA, and signaling taking place among our millions of cells, it seems silly that we might make a decision based on a single measure, such as a lab value measured at one point in time, with little additional information about context or information relevant to activities taking place in our daily lives. Of course, what he neglected to mention (or perhaps understand), is that these complex systems have evolved over millions of years to function without human analysis or consciousness, and that one of the first rules any physician learns is that you can't outsmart nature. Nonetheless, sometimes even these highly tuned systems can break down requiring us as patients and providers to attempt to 'open the hood' and figure out what went wrong and how we can fix it. In this respect, he's entirely accurate that it would be much better if we did have access to at least some of the indicators of this highly refined system in order to mend, or better yet prevent, further damage to our bodies. And he's also correct that it is unfortunate that with all of the advances in engineering for collection of detailed, real-time biological information, we have made few inroads into getting this information into the hands of experts or systems that might use it to improve our health.
In this section, we're going to dive into the issue of data collection for the individualized approach. We will start with a general review of three key principles guiding collection of data, and then discuss in more detail some of the sources of data that we can access to better understand and quantify our health and wellness, starting from the least invasive and moving inward toward the most invasive.
Principles of Individualized Data Collection
1) Nature of the Data collected
Orthogonality is a mathematical term that refers to the degree to which two elements in vector space are situated perpendicularly to each other. In being perpendicular, there is no correlation between these elements, and therefore there is no overlap in the information that each is providing. In thinking about the type of data that we might want to collect about an individual person, orthogonal information from two or more sources, because they are not correlated, will provide the most 'bang for the buck' in terms of providing the most information about that person. In contrast, two or more data sources that are highly correlated will not provide as much information since one source can be derived or mapped to the other. For example, if I choose to analyze data from two monitors, one a wearable heart rate monitor and the other an implanted electrocardiographic device (see below), both of which are providing a measure of average heart rate, I'm not really gaining much additional information from using two monitors since they are supplying the same information (resting heart rate) for my analysis.
On the flip side, two or more highly correlated sources of data can also be useful in situations in which the integrity or accuracy of one source is in question, in which case, the correlated data sources can be used to corroborate the accuracy of each other. In our example above, the wearable heart rate monitor may detect my blood pulse for each heart beat, but suffer from excess noise due to motion of my hand during activities. The implantable electrocardiographic monitor may not have as much noise from activity, but may not accurately reflect the blood flow (pulse) from each electrical beat if I have electrical impulses that do not cause blood to flow. If I were interested in performing an analysis based on an accurate measure of my pulse (blood flow), then I might opt to use both as they can corroborate each other due to the correlation in what they are measuring.
Now, important in this discussion is the actual monitor or source of data being provided. Implicit in the above discussion is that the actual source of the data is from two separate, independent, monitors. Many wearable monitors, such as Fitbit devices, are only providing data from one source, such as an accelerometer, with the different outputs (calories, steps taken, distance) being not only correlated, but completely derived from the signal from that one source. In this case, calories and steps would not be useful for corroborating each other since, due to being from the same source (the accelerometer), they will likely suffer from the same inaccuracies in measurement (noise, etc.). In this case, it is unlikely that conducting an analysis with both of these data inputs will have any greater accuracy than including only one. [Note: this discussion assumes that the data being used is being collected simultaneously. In contrast, sleep quantity can also be provided from the accelerometer on a wearable device, and might be useful if combined with steps since it is not being collected at the same time. In this case, it would be providing orthogonal information to steps or activity.]
To summarize, data sources should be independent whenever possible, and should be assessed for whether they are providing orthogonal information or corroborative information. Failure to understand how they are combining can lead to practical as well as statistical problems in our analysis (see multicollinearity for more details).
2) Practicality vs. accuracy
There's a famous parable in research called the drunkard's search (also called the streetlight effect) that goes as follows:
A policeman sees a drunk man wandering around under a streetlight searching the ground late one night. The policeman asks him what he's doing, and the drunk replies that he's looking for his keys. After searching for some time together, the policeman asks him if he's sure he lost them here, to which the drunk replies that he probably lost them back along the dark street somewhere. The policeman then asks, incredulously, why he is looking for them here, and the drunk replies, "because this is where the light is."
In an ideal world, we would be able to collect information about any biological or lifestyle marker we want in order to understand the biology of that person and to improve their health. Unfortunately, people need to be able to live their lives, and the only good method to collect most of this information is through highly invasive methods. Practicality almost always trumps accuracy in medical research, and while we hope that the information we are able to collect will provide more insight than the drunk searching for his keys example above, it is not surprising that much of what we know to date about medical risk factors comes from settings in which data can be easily obtained.
High blood pressure measured at an office visit is a well-known risk factor for future heart attacks. However, high blood pressure is really only indirectly linked to the pathophysiological process of having a heart attack. It is only in extremely rare instances that high blood pressure can directly cause a heart attack; in most cases, it is simply a marker that research studies measuring various parameters during a doctor visit have found to correlate with an increased risk of a heart attack in the future. It is very likely that there are many more proximal signals (biomarkers, etc.) for an impending heart attack in a person's body, which if we could only measure them easily, would render our use of blood pressure obsolete. However, it turns out that people aren't too excited about having their blood drawn at regular intervals throughout the day, over the course of their lives, and thus we have no choice but to use the most practical measure, blood pressure at an office visit. [There are emerging studies of so-called ambulatory blood pressure, checked throughout a day, but none have shown a directly causative proximal effect on heart attacks to date].
Practicality of data collection is a very important principle in monitoring, and one which should be considered in any future ideas about sources of information, although unfortunately the practical aspect is too often overlooked. Implicit in this principle is the contrast in active and passive data collection, in which data collection approaches that require active collection methods, ranging from daily dietary journals to blood draws to manual blood pressure measurements, generally are not reliable for consistent data collection over time. Passive collection methods, such as through wearable or implantable monitors or through data mining from grocery store purchase records or hospital visits, are much more reliable for moderate to long term data collection. We will discuss some of these in detail below. Analysis of individualized data requires large amounts of longitudinal data collection in order to overcome the issues of seasonality or to identify trends in data, and for that reason, we should almost always aim for the most passive approaches over active data collection whenever possible.
In summary, practicality of data collection is a key principle to consider in selecting or designing monitors or sensors, with passive data collection being the gold standard for approaches. Advances in technology are slowly bridging the gap in terms of how much data can be collected passively, and as we will discuss below, there are reasons to be optimistic that we can do much better than single blood pressure measurements in a doctor's office a few times a year in order to identify individualized predictors.
3) Privacy and Safety
It is fitting that we follow-up a discussion about passive data collection with one about another important principle in data collection: privacy. The downside of collecting a lot of data about someone without them needing to alter their lifestyle patterns is that we may collect data at a time when they preferred we did not. Whether we're mining shopping data or analyzing data from an activity monitor, there is a risk that we may find a certain purchase that could implicate the patient, or identify activity during a time when the patient was doing something they preferred be kept secret. The upside of active data collection is that, for the most part, the subject has total control over what is collected and when. Patients do not always tell their doctors everything, and we all have things that happen in our lives of which we would prefer no one had knowledge (much less a permanent or searchable record). Passive data on internet browsing patterns, social media posts, and driving routes has been collected for many years by companies for monetary purposes, and only recently have we begun to understand the implications (although institutions are now recognizing the situation, and taking steps to prevent loss of privacy).
Privacy of healthcare data has been of well-recognized importance for some time now (see HIPAA for more information about this issue in the United States), although the lines can blur when we move from healthcare uses to so-called 'wellness' applications, such as wearable activity monitors. To say that it is a moving target is an understatement, and it is likely that over time we may have a new set of regulations about what kinds of lifestyle information can be collected, stored, and shared, even if it is being used specifically for healthcare. Further, the role of apps and consumer-directed information technology methods in healthcare is increasingly being recognized by the FDA and other administrative bodies as an area where they may need to step in to prevent abuse or hard to patients (We will address this situation in more detail in Section #2 about User Interfaces).
Regulatory issues aside, there are safety considerations for any sensor or monitor that should be addressed before use in individualized data collection. Implantable devices can become infected, which at a minimum can require emergent surgical removal, and in extreme situations can result in bacteria seeding the blood stream causing system-wide infection (called sepsis) and damage to heart values or seeding of prosthetics requiring antibiotics and/or surgery, if severe. As in routine healthcare, it is important to weigh the risks and benefits of implanting any monitor in the body, and equally important to have close follow-up to ensure that devices are functioning appropriately and not causing adverse effects. In Section #3, we will discuss the role of clinicians in the use of sensors and monitors in more detail.
To summarize, no data collection approach is completely risk free, and we should consider risks, including privacy and safety, when choosing any system. Although many monitors, including implantable sensors, are generally safe for use, we should ensure that patients understand the risks before application, and (if needed) ensure that the appropriate governing body (i.e., FDA) has approved use of a monitor in people.
Types of Data Collection Approaches
Keeping the above principles in mind, we will now focus on some specific approaches to data collection, starting with the least invasive and moving to the most invasive (i.e., implantable). We will provide a few examples, although keep in mind that new approaches and devices are emerging every day, and this will not be an exhaustive list of all possible approaches. Like many of these sections, we anticipate that there will be significant overlap with other sections, especially Section #2, since many data collection approaches can be built into a smartphone app, as well as Section #3, since some data collection approaches require intensive interaction with providers.
1) External Data Sources
Data mining of external data sources is among the least well-understood by many medical investigators and providers. Although recent publications about use of Google Trends and other repositories of external data for social science and economic investigations have enlightened some of the power of these data sources, they remain poorly utilized in medical research. Among the reasons for this lack of understanding is that the 'backwards' approaches for data ascertainment applied in accessing this type of data (i.e., data mining, scripting, database queries) are the exact opposite to how most clinical researches collect data for studies. The conventional approach to a clinical research study is to create an empty database of every parameter that the investigator thinks could be important, and then to populate it by using chart review, patient interviews, or clinic measurements until a sufficient number of patients have been 'enrolled' as to provide statistical power. [Note: the approach described is for retrospective clinical studies. Prospective studies, such as randomized controlled studies, also involve 'forward' data collection, although in those cases the reasons are obvious].
However, not only can more information be obtained by applying a script or data mining technique to a large, unannotated, collection of data, but the specific criteria used to include or exclude a piece of information are written explicitly into the script, and can thus be audited by a reviewer seeking to examine or replicate the methods used. In addition, and relevant to the discussion on passive data collection above, these approaches can be applied to data collected for non-medical reasons, but which might contain information that could provide a more detailed picture of an individual's lifestyle and risk factors for disease. Although these approaches are only starting to be employed in research, below we will highlight a few potential sources of individual-level data, and some methods through which it could be applied.
1) Grocery store purchase records. Most large commercial grocery stores have realized that by incentivizing customers with automatic rebates and discounts, including gas, they can coerce them into using a 'loyalty' card that keeps track of all purchases at that store. Since many people tend to shop at a single location for much of their food, this information could potentially provide important information about dietary patterns and changes over a period of time. Of course, there are caviats, such as whether a person lives alone or in a large family, and whether he or she uses alternative sources for certain foods items, but keep in mind that we're looking at individualized prediction models, in which any source can absolutely be useful for one person, even if on average across a population it provides more noise than signal.
How might grocery store purchase records be used to collect individual data? First, the patient provides access to the full, time-stamped, record of purchase items for the investigator over a period of time. The investigator then parses the record in order to categorize the purchases by type and relevance. Obviously, cleaning supplies or dog food could probably be discarded, while vegatables or meat purchases might be very useful. After parsing the data, each category is then binned for each time point, and a multivariate time series of food purchase categories is created. At this point, a given outcome is included in the model, for example weight. A model is then created in which purchase (and inferred intake) of each category is modeled against daily weight in order to identify correlations. If vegetable purchases are noted to correlate inversely with weight, then the patient could be advised to each more vegetables. If no category is correlated, then the data source could be discarded. Again, it doesn't necessarily need to be predictive in everyone. (See Section #3 for a full example of clinical application of the individualized approach).
2) Medical Record Data. In many ways, mining medical record data seems like the most straightforward way to obtain useful individualized prediction information, and in an ideal world, it would be (see above about practicality). However, there are several reasons to be cautious about relying specifically on medical record information when attempting to draw inferences about individualized outcomes. To understand the challenges, it will be helpful to dive into what exactly is contained in the medical record (also called the electronic health record, or EHR). The medical record is a complex collection of notes by physicians, proceduralists, nurses, administrators, social workers, and office assistants; lab values and results from medical studies, such as x-rays and MRI's in various degrees of accessibility (too many in PDF form that makes data mining extremely difficult); medication lists of varying accuracy; orders for future tests, referrals, or procedures; and billing codes containing diagnoses and procedure codes. The accessibility of the data is highly variable across healthcare systems, with lab values and billing codes tending to be the most accessible, while office notes and test results (other than labs) tending to be among the most challenging to access information systematically. There are advances in text recognition and natural language processing that hold great promise to expand what is available in the medical record, although for the most part these are not widely available.
The most widely used data from the medical record are billing codes, which at first glance would seem to be an excellent source of information for individualized medicine. However, keep in mind that the incentive for coding diagnoses by a physician is highly monetized, and most practicing physicians are all too aware that certain codes can bring in far more money than other codes, even if the diagnostic difference between the two is in the grey zone. In addition, physicians tend to code only for those conditions that they treat, which can be a particular problem if a patient sees a specialist for one condition, and a primary care doctor in a different health system (and different EHR) for others. The former may exclude 'irrelevant' diagnoses, even if these are going to be relevant to the condition you would like to study. Finally, even the most high-maintenance patient will only tend to see his or her doctor a few times a year, and so the granularity of information collected from medical record data is unlikely to reach the levels needed for individual analysis (see Section # 4 on Analysis). In many cases, if documentation of a recurrent medical condition, such as headaches, is desired, this can be obtained directly from the patient, without needing to go through the process of obtaining medical record data (we're leaving out the bureaucratic issues of obtaining permission to release data to patients or external providers).
3) Social media posts. Collecting and storing social media data for research is a bit of a controversial issue these days in the setting of companies like Cambridge Analytica; however, if you can see beyond the lines of a few bad actors, there is immense potential information available in social media data. For many people curating a Facebook or Instagram page, or posting regularly on Twitter, much information can be gleaned about a variety of potentially medically related issues. Psychological conditions, such as mood or anger, can be ascertained, as can time spent with family, traveling, or going out with friends. In similar fashion to how we parsed and then categorized grocery store information above, social media information can be used to model all kinds of lifestyle and health correlations. In addition, it can provide valuable outcome data, particularly if patients post information about recurrence of certain disease conditions. Of course, the privacy issues are no different than any potentially sensitive source of information, and especially assurance about appropriate data custody and sharing would need to be cleared before using the information. However, there are many who believe that social media holds great power as a big data source for medical investigation, and it will be interesting to see what creative methods are applied to this data in the future.
2) Logs and journals
We discussed above the challenge of active data collection, in the form of diet logs or symptom journals. However, if a person is able to collect this information in a systematic and regular method, it can be highly useful for individualized data analysis. There are a few strategies that can be applied toward a goal of regular data collection, and we will highlight a couple here. In Section #2 on User Interface, we will also discuss how app design can improve this process.
The first key to regular active data collection is making it a habit. We all have habits, and one of the best ways to ensure that a certain task is accomplished regularly is to turn it into a habit in which one doesn't need to think before doing it. Clicking the seatbelt in a car is a habit that most people have been able to work into their regular driving routines. In the hospital, many in our generation of physicians have developed a habit of pumping hand sanitizer before and after entering a patient's room. These are not activities that require great thought, and thus it is not clear whether active data entry could be made as thoughtless and routine as clicking a seatbelt; however, this is a good goal for implementing active data collection. To help in the process, another tip I've found useful is to find a way to include data entry into some part of your everyday life. I keep my scale by the shower, and weigh myself each time and record the value on my phone before getting in. Others have found reward as a useful incentive. One patient would not each dessert each night until she had logged her meals into a dietary log after dinner.
Technology can be another useful method for routine active data entry. A reminder on a smartphone to enter data is one approach, or better yet, programming your Google Home, or similar device, to prompt you at a time when you're in the kitchen and can enter data. If there were a way to enter data entirely without writing anything (tell Google Home what you ate that day), then you wouldn't even need to write anything down. We have developed an iPhone app (PM App Lite) that includes two options for entering data. The first is the routine daily log of data entry, and a second is a symptom-only option, so that a person only needs to remember to enter data on a day when he or she has the symptom or interest. The analysis then adjusts days without data entered to reflect a lack of symptoms on those days, and is able to create a model of disease recurrence. Gamification or using social pressure (competition between friends) are other strategies. Like the use of social media as a data source, it will be exciting to see what ideas arise to incentivize people to log data regularly down the road.
3) Wearable monitors
It would be quite the understatement to say that the market has exploded with wearable monitors, particularly activity monitors, such as Fitbit or Garmin. These devices, which have been targeted toward the 'health-conscious' population as a method to keep track of daily activity and exercise, as well as to set goals and compare with friends and family, are perhaps the most obvious demonstration that, if appropriately incentivized, people will allow nearly continuous monitoring of their personal lives. Of course, not everyone has a Fitbit, Apple Watch, Garmin, or other wearable activity monitor. For some, the reason is economical (keeping track of steps every day is a First-world concern that people who have to worry about where the next meal will come from frankly don't have time to even consider). For others, convenience is a factor, as these watches, in contrast with your standard Timex, require frequent charging, and often require syncing with a smartphone or other device in order to make adjustments, like changing the time or date. Even among users, there is hardly a well-defined medical reason for use. Despite the 'goal' that Fitbit sets for each of us to reach 10,000 steps per day, there is little to no clinical evidence to suggest this number, or frankly any other daily step value, has any important prognostication for disease avoidance beyond what the majority of the various medical societies (i.e., American Heart Association) recommend for daily activity and exercise. Nonetheless, many people do choose to wear these monitors, and attempt to reach the activity goals, and compare with friends and family. As a result, these monitors provide an opportunity for investigators and clinicians to examine the data and make inferences about how it could be used to improve health.
There are practical differences between the brands of activity monitor, or other wearable monitors, when it comes to accessing data for analysis. Fitbit data, which any readers of this site can attest, is quite available to users and third parties (with user permission) in a number of manners, from CSV files that a user can directly download from the Fitbit website to an easy-access API that allows apps to access the data. Our group has used the latter approach to develop an app that will perform a range of individualized analyses, as well as allow comparison with friends, family, etc., called the Udatalyzer (udatalyze.com). Apple Watch, which also allows access to data, albeit via the Apple development platform rather than a simple API, is another fairly developer-friendly wearable for investigators (particularly those located near Cupertino, CA or the Stanford University campus). Garmin has an API, but requires a fee for developers to access the data, and thus our preference has been to focus on the others.
It will be interesting to see where the world of wearables moves over time. There is definitely an interest by companies such as Fitbit to break into the medical world, and with it the increased demands of FDA approval, etc. On the other side, medical device companies have been toying with the idea of moving into the commercial (i.e., wellness) world of wearable monitors, although they have not made the move to date. Anecdotally, it is not unusual for me to have patients asking me to interpret the heart rate measurements from their Apple Watch from a medical perspective in an attempt to correlate with a clinical condition. Many have told me that their first indication of a heart rhythm problem came from a reading from a Fitbit or Garmin. Of course, I have to remind people that these devices are not designed or FDA-approved for medical uses, although it rarely stops them. It's hard to blame the empowered patients because most device companies release little to no data back to them directly, and patients are often at their doctor's discretion about getting data from a monitor, even one implanted in their own body (see below).
4) Biosensors and noninvasive biological sample collection
This category is perhaps the most exciting in terms of potential future sources of useful data to guide individualized management of medical conditions. Many of these sensors did not exist even ten years ago, and it is likely that many will be obsolete within years as technology improves. Broadly, this category can probably be broken into several smaller categories based on whether the devices are currently in use, collecting biology information or actual samples, or are in the pipeline for future uses. We'll present examples of all three.
A Holter monitor is a heart monitor that records the electrical activity of the heart continuously over a period of anywhere from 24 hours to 2 weeks. Holter monitors have been around for many years, although the key challenge has been that the monitors are bulky, and require the patient to be connected to wires that can disrupt activity and sleep. However, recently a company called iRhythm has developed a heart monitor that is little more than a sticky patch, which the patient attaches to his or her chest and which stays in place for up to 2 weeks recording information. This Zio patch, is waterproof, and other than occasional skin irritation, has minimal interference with daily activities like older versions. The Zio patch has spawned many imitators, and it is likely that these monitors will only get smaller and with longer lifetimes over time.
The Zio patch collects electrical information about the heart, but there are currently methods to sample biological information as well. One excellent example is glucose monitor technology. Historically, a diabetic needed to prick his or her finger several times a day, before meals, in order to measure the blood glucose level so that he or she could figure out how much insulin to take. As you can imagine, this was uncomfortable and inconvenient. More recently, glucose monitoring technology has advanced from some non-invasive methods, to other continuous glucose monitors that essentially act like an artificial pancreas (the organ that regulates blood glucose). Investigators have already began studying this data to identify associations with health and wellness even beyond diabetes. It is likely that similar approaches will be developed to measure other biomarkers circulating in the blood stream, which will open the door to all kinds of potential insights into disease risk and health. Currently, most biomarkers that are examined from a patient's bloodstream are measured once or twice in a year. One can only imagine how much we could learn from monitoring these biomarkers multiple times a day.
You probably wouldn't think of your smartphone as a biosensor, but investigators have figured out a number of ways to extract potentially useful information from the accelerometer and camera in smartphones. The accelerometer is obviously used to track activity, but researchers have also recently used this sensor to monitor tremors in patients who suffer from conditions such as Parkinson's disease, for which tremor is a major comorbidity. This information can be used to guide and adjust treatments, with potential to be far more useful than a simple biyearly office exam. Facial recognition software is another opportunity to collect longitudinal data about a patient's mood, which again can be used to guide treatment and monitor symptoms of depression or bipolar disease. This information could go much further than current office interviews to obtain objective information about a patient's mood trends, likely capturing information that would have been missed or unnoticed by asking a patient how he or she felt.
5) Implantable Monitors
Our final category moves into the realm of the fully invasive approach to obtaining individualized data, with implantable monitors. As discussed above, it is important in making the choice to implant a monitor that the patient and provider have weighed the risks against the benefit, and that the patient is aware of the potential risks with having a monitor in his or her body that cannot be easily turned off or removed. That being said, implantable technology may very well be the future for individualized medicine as it virtually guarantees long-term passive data collection for which a patient doesn't need to remember to put on the wearable device, or charge it, or avoid smashing it during an activity. There is great power in the amount of data that could be collected, and like many implanted devices we use medically already, such as pacemakers, patients eventually don't even notice that they are there.
There are two heart monitors that are implanted under the skin and allow for up to 3 years of uninterrupted continuous monitoring of heart rate, as well as activity (and likely other sensors, although the companies have been hush-hush about them). The Linq monitor is made by Medtronic, and the Confirm is made by Abbott (formerly St. Jude). The Confirm connects via bluetooth to a smartphone, although to date Abbott has not allowed patients to directly access the information from it. In fact, this is arguably the biggest drawback to using these monitors for individualized data collection beyond the infection risk, which is that many medical device companies tend to want the data only for themselves, with patients and investigators often needing to jump through multiple hoops if they want to obtain the data in any format other than what the companies choose (typically a PDF, which is basically useless for analysis). Our own experience attempting to get medical data from medical device companies for analysis has been challenging at best, but hopefully these companies will gradually change their minds about collaboration with greater appreciation for the role of individualized analysis of the data produced by their products.
Collaborative potential aside, it will also be interesting to see what direction things go in the future concerning implantable devices for individualized data analysis. On the one hand, the potential for passive data collection without interruption seems ideal over devices or monitors that are external to the body. On the other hand, advances in technology for wearable or external monitors may obviate the need to implant, with imputation or other techniques making up for the interruptions in data collection. One thing is certain, the future is definitely bright for more and more data sources for individualized data collection, which will shift the burden from the engineers designing these devices to the analysts and clinicians, who must figure out what to do with the data. This is the subject of future sections...
Next, Part II. User Interface