Baymax and Scarlett: Predictive Analytics and Machine Learning (Data Part II)

Last week  I encouraged Patient Support and Patient Market leaders to evaluate their post-launch data needs at least one full year before their product’s PDUFA date. The usefulness of data is not limited to historical program performance review. As I wrote earlier,

You do it to identify actionable insights and give your colleagues the tools they need to deliver on your brand’s plan of action (POA). 

Analytics, specifically predictive analytics and artificial intelligence, can be powerful tools that your team can leverage to reach and help patients. Humor me while I try to wield my Mama superpowers to explain the difference between predictive analytics and artificial intelligence.

Any maternal stance I had on “screen time” went out the door during COVID. Monster Jam (and I mean, ALL of Monster Jam… the trucks, the videos, the playsets– including those with that horrid kinetic sand) saved my family, as did television. LOTS of television. Not even good tv. Our son watched a lot of kids on YouTube hawking toys.

BUT we discovered Big Hero 6 and Baymax. The 2014 winner for Best Animated Feature at both the Golden Globes and the Oscars, Big Hero 6 is a completely charming movie. Sure, there are some dark themes (why does Disney have to kill off parents?!?), but there is also some very thoughtful discussion on managing grief and allowing friends to help you. The Bala grown-ups also loved that Big Hero 6 makes being smart, going to college, and studying science super cool. And that’s in no small part thanks to Baymax, the inflatable, marshmallow-y robot who serves as a “healthcare companion.” Baymax is part nurse, part lab diagnostics, part therapist, and deeply committed to his patient’s best interests. 

As lovable as Baymax is, however, I wouldn’t describe him as intelligent. He analyzes all data inputs against his vast warehouse of knowledge in real-time, with incredible speed. BUT, his responses are limited to what he has been programmed to do. If Baymax does not have a pre-coded reaction corresponding to a given input, he does nothing. Baymax can only do what the Hamada brothers, Tadashi and Hiro, have programmed or instructed him to do.

Not for nothing, this sing-songy fist bump is EPIC among the preschool set.


Baymax is like your predictive analytics platform: decisive action based on data, but that action is limited to the direction given. 

Here’s an example: 

  • Oral medication
  • Taken once/day
  • Dispensed in 30-day quantity
  • Distributed via Specialty Pharmacy

Let’s say that de-identified, aggregated patient data review indicates:

  • Commercially insured patients who:
    • Could be eligible (source: SP claims data) for manufacturer copay support
    • But are not enrolled in copay support (copay program enrollment)
    • Are more likely to fall off therapy between prescription fills #2 and #3 (SP dispense file).

Your Patient Engagement team creates a campaign with patient messaging specific to different milestones on the treatment journey, including information about copay support enrollment (automatically sent on days 15, 35, and 42 post-initial prescription fill).

See what we did there? We used a data set for:

  • Analysis: We identified patterns in patient behavior.
  • Strategy: We leveraged our understanding of our patient (clinical trials, market research) and coupled that with the data to: 
    • Explore the patient motivations behind the behaviors captured in the data.
    • Devise a data-driven plan to encourage positive patient behaviors.
  • Execution: We will communicate with these patients via the channel most appropriate for them: in this case, SP-initiated message sent via SMS text. True, the manufacturer will pay an enhanced services fee for the SP to send these messages. Still, there may be no other opportunity to capture these patients if they haven’t opted into other manufacturer support programs. And because it is worth repeating: at no time should the manufacturer have patient PHI.

The strategy and execution outlined above hinge not upon data, but rather the human interpretation of the patient behaviors behind that data. The Patient Engagement team makes the rules. They are Hiro, so to speak.

Let’s talk about a subset of Artificial Intelligence called Machine Learning, and another Academy Award Winner (Best Original Screenplay) decidedly NOT for kids: Her. Written and directed by Spike Jonze, Her features Joaquin Phoenix, Amy Adams, and Scarlett Johansson’s Voice. 

Phoenix’s character Theodore Twombley falls in love with an operating system (OS), Samantha, played by Scarlett Johansson. Theodore wears an earpiece and microphone, and Samantha “sees” via his phone’s camera. Everything he says, hears, and sees… Theodore’s data, so to speak, is captured and analyzed by Samantha. Unlike Baymax, however, Samantha can learn. More specifically, she can extrapolate, anticipate, and create.

How could machine learning change the earlier product scenario? An AI platform could

  • Extrapolate: Through fast, iterative processing of claims data, determine which patients are likely to have health plans that could authorize 90-day fills. 
  • Anticipate: Estimate how much is remaining on all potentially eligible patients’ annual deductibles and where the copay program stands overall in utilization and total benefit spend. 
  • Create: Automatically generate a message to both the patient and the prescribing HCP about switching the Rx to 90-days precisely 17 days after the initial fill. Additionally, the patient does (or does not) receive messaging regarding copay enrollment, depending on the patient’s need, the likelihood that copay support would positively impact patient behavior, and the program’s financial health

Wait a minute, that had nothing to do with the previous example on copay. Or did it? In the earlier example, the marketing team devised tactics to ensure adherence through at least the first four prescription fills. The second example encourages medication adherence through at least four fills.

Which approach might offer a better return on objective?

If you’re expanding your brand’s current analytics and AI utilization, don’t be afraid to ask the following of your potential vendor partners:

  • What will they need? Time? Patients? Transactions per patient? How many must be captured before their platform can provide meaningful insights? Brands often “hold off” on digital marketing for at least one-year post-launch. The problem with this approach is that you may have missed the opportunity to capture the data you need to drive your marketing campaign.

Your insights will only be as good as your data inputs.

  • For new first-in-class therapies, where do they start? Can they give examples of how they have worked with purchased data sets?  What do they recommend for identifying possible product analogs? Therapeutic area, prescriber profile, anticipated coverage, channel distribution, list price… the product or therapeutic area that you think most closely mirrors your patient experience may be wrong.
  • Do they understand patient data privacy and the unique regulatory challenges in manufacturer Patient Support Programs? Can they explain their logarithms to your Legal and Compliance teams? The manufacturer’s brand team isn’t trying to sniff out a vendor’s proprietary technology. But there will, without fail, be questions about how certain patient profiles are developed and used.

Just as spreadsheets transformed accounting and corporate finance, machine learning has the potential to transform Patient Support and Patient Engagement Marketing. Eyes can gloss over as soon as the word “data” is uttered. “Artificial intelligence” suggests cold, unthinking machinery.

But the truth may be that machine learning enables us to personalize access support programs. Don’t we want to communicate with patients in ways that resonate with them and who they are at that moment in time?