There’s naturally a lot of buzz over the acquisition of One Medical by Amazon.com. All the big tech companies are convinced they can fix our health care system. Perhaps they will start serving up appropriate treatments the way they choose the movies we view or the political diatribes we consume. But I recently talked to Gus Malezis, CEO of the digital identity company Imprivata, to explore what Amazon has to do to make its investment pay off.
Malezis pointed out that healthcare landscape is strewn with high-tech failures such as IBM Watson and Google Health. Patient data is just too scattered, too inaccurate, and too complex to yield actionable insights at this point.
Artificial intelligence (AI) is powerful and will pay off eventually. Malezis and I talked about the steps that could lead to success in healthcare. The tasks he listed will be familiar to most readers of this web site. Let’s delve into the challenges involved.
Step 1: Identify the patient securely
One could well guess this to be the first concern for the CEO of a company like Imprivata. Most clinicians don’t make patients go through a validation check like a financial institution does, or even a government office issuing an ID.
Their reluctance to make the patient undergo such a check is understandable: It’s a hassle for both clinician and patient. Experian can reject you because you got a letter wrong when giving them your previous postal address. And when I most recently renewed my driver’s license, I found that the RMV put up three web sites with three conflicting lists of what identification to provide.
But if you want good data, you need validated identity. You’ll probably have to enroll patients without initially requiring that step, because other clinical institutions don’t require it and you want to make onboarding easy. But at some point—hopefully before you have to deal with a health emergency—it’s worth trying to validate each patient.
Step 2: Gather and harmonize as much patient data as possible
If identifying a patient securely is hard, the task of retrieving their data from other sites is truly intimidating. The patient might have been at dozens of different doctors over the course of their life. Luckily, the most important information about their conditions has been collected by each clinician, so it should be enough to retrieve records from a few current and recent visits.
Still, this is another expensive investment for both your data-driven practice and the patient, who has to sign release forms and perhaps transport the data to you.
Clinicians are fiercely protective of data on patients (even though the information should really belong to the patient) and drag their feet on releasing it. Outmoded tools that make it hard to collect and release data give the overworked clinical staff even more reason to resist. Some institutions still hand you a pile of printed paper when you request records, even when those records are in an electronic system.
The lack of a universal patient identity, as described in the previous section, makes it notoriously hard to know whether you’re getting data on the right patient. Clinical settings traditionally apply a wide range of checks (name, birthdate, address, etc.), but each is fuzzy given the lives patients lead.
Assuming you have a large data set on the patient, the data has to be harmonized—probably the biggest barrier to deriving useful insights through AI and analytics. The problem is not just that data is in different formats; software tools can overcome such mismatches. The problem lies more in the inconsistencies of diagnosis and treatment.
Some doctors have favorite codes, leading to different doctors assigning different codes to patients who have the same essential conditions. Upcoding and other billing tricks introduce inaccuracies. Doctors may also say different things about the context of the diagnosis.
The healthcare field’s understanding of conditions also changes over time. For instance, psychiatrists are diagnosing a lot of people with bipolar disorder nowadays. These patients might have received other diagnoses, such as schizophrenia or borderline personality disorder, ten or twenty years ago. Such changes muddy the records.
Finally, the same condition could also be treated by a variety of medications, and it’s hard to find digital libraries that indicate which medications are interchangeable for analytic purposes.
Anyone who has worked on a data analytics project knows that each of the problems cited in this section could ruin a project. Malezis says that data gathering should be incremental, given the barriers and expense involved.
Step 3: Evaluate the patient’s current health status
Clinicians don’t always agree on conditions, and prefer to conduct their own examinations. Many specialists require a test even if the patient got the same test a month or a week ago at a primary care practice or another specialist; some of this is necessary, but adds to health care costs and consumes time. Malezis thinks that any organization taking on responsibility for a long-term relationship should ask the patient to undergo a physical, to ascertain the current state of health, and review the consolidated patient record to understand the patient’s health history. This baseline of medical history and current context will be of incalculable value in an urgent situation when time is restricted and timely response and treatment are key.
Step 4: Analyze the data to determine a treatment plan
Now, hopefully, you are ready to use AI to determine what to do with a patient. But again, the trajectories of IBM Watson and Google Health are warning signs. The previous three steps are necessary to give you good data, and hopefully accurate models.
Putting all these challenges together, Malezis estimates that producing useful models for treating patients will take Amazon no less than five years; likely even more for comprehensive and helpful output. We’ll be watching the big companies that are battling over digital health during that time.
