Data Rich Innnovation Poor
Pharma Marketing Was
In 2020 I started working for the first time in the Pharmaceuticals Space. Coming from CPG it was an interesting challenge to understand the dual customer dynamic of HCPs and Patients. The industry is immensely data rich, with details on the who, how, why, and when of a product interaction easily accessible. The irony though is that the Pharma space while very data rich did not really leverage this data in any particularly advanced ways.
Much of my time was spent with clients discussing foundational Marketing Science Principles like Taxonomies, Measurement
Frameworks, and Learning Planning, but there was particular area of pharma marketing I really wanted to change, segmentation.
Perscription Weighting
Defined Communication
Parama was going through a period of rapid modernization in 2020,
as it looked to catch up with many of the data driven marketing innovations prevalent in CPG, but a lot of their core segmentation principles were founded in a legacy process known as Deciling.
This process ranked HCPs based on their prescription volume, and grouped them into deciles where top prescribers would be in the
10th decile, and those with the lowest activity fall into the 1st decile.
A simple way of defining who their important customers were, but
one that really failed to take into consideration any trending within
the group, and imposed a limit on how a customer could be engaged.
Visibility on Potential
There was Limited
The problem with segmenting your customers based purely on the amount they are currently prescribing is that you have no means of understanding whether that customer was always a high priority or had recently become one. Limiting your ability to communicate effectively with HCPs who may still be prescribing small amounts but have the potential to increase their prescribing weights. This is also true of HCPs who may be starting to move their patients to an alternative therapy, there is no trending within deciling so you have no way of knowing whether the current.
​
The use of Rigid Deciling was inhibiting our ability to cater
communication. HCPs that were moving to new therapies were receiving the same communication as those growing. And those growing were considered ignorable until such time as they started weighting more heavily. We needed a better approach to prescription weighting, one that took into account an additionalvector.
Machine Learning
and Chronovectors
We needed a means to cluster based on trending over time, but incorporating this new vector into a clustering calculation was pretty challenging using traditional methods. Processing the calculations necessary to do clustering over time and aggregating them into targetable groups would require time and resources that would make it cost prohibitive.
The solution came from an unlikely source, Netflix, or more precisely the Machine Learning algorithms Netflix uses to understand viewing behaviors and serve their users suggestions. In 2021 the great AI craze had yet to hit full gear, but we had been leveraging Machine Learning for heavy data processing for some time. In simple terms, once a model is trained, it allows us to do calculations that would take an individual hundreds if not thousands of hours in an a fraction of the time.
Funding, and Finding
The Perfect Use Case
What we had was an Idea, and the algorithm and process coding required to make it a reality, but what we needed was funding. In 2021 the demand for AI is not what it is today, in fact the general level of understanding for what Machine Learning can offer was quite low. We were also using the algorithm in a way that had never been done before. We needed to convince a client to invest in creating something completely new that wasn’t 100% guaranteed to succeed.
The product could conceivably provide benefits to any segmentation and communication plan, but we needed a client that would see significant marketing impact when enacting this capability. This is
where Takeda stepped in, they had a product Trintellix that was experiencing declines thanks to emerging competitors. When a HCP switches therapies the results are rarely immediate, discerning between a natural month over month decline and those that are
starting new patients on new therapies is something that our algorithm could solve for far more effectively than any equation based solution. Trintellix was the perfect use case, and if we got it right could effect significant business impact.
Training and Validating
We had our product plan, and our funding partner, the next step was
training. Through our funding partner we now had the means to acquire the data needed to train the model. Not just on Trintellix prescription data but on a whole host of Therapy prescribing patterns. The more data we had the quicker and more viable the clustering segments we produced would be. Fortunately the process was very smooth, and within 4 weeks we had trained a model that could produce 16 groups across 4 tiers of clusters with aligned aggregated trending and the corresponding confidence Interval of each HCPs association with said trend line.
We took 250 HCPs that fell within a 95% confidence of the assigned cluster and manually checked to see if the assignation was indeed correct. Despite the confidence limit we found that 248 of these did indeed fall within said cluster. We now had a way of finding declining HCPs more effectively than any existing segmentation process.
Translating the Results
Into Marketing
The extracts from our algorithm would not make sense to anyone who isn’t intimately familiar with the process. To make the results more impactful we distilled the plot outputs into 3 categories, Growing, Steady, and Decline. We subsequently were able to create clustering groups of High, Medium, Low, and Base tier prescribers and assign their generalized trend.
This allowed us to create targeting groups that we could address directly on their recent decisions to move away from Trintellix as a therapy. Our Reps were now armed with information to guide the conversation to the reasons why heavy prescribers were moving away, our e-mail and media teams could actively target mid and low tier prescriber to ensure Trintellix was back in these HCPs top of mind. More than just being able to target effectively it meant we could be more efficient in our Marketing, allowing us to balance more effectively against over exposure to already growing prescribers towards those where we were losing ground. The end result was our test case in Trintellix was able to halt their recent quarterly declines and actually started achieving growth again.
Using Machine Learning For
Acitonable Marketing Impact
This would be Klick Health's first machine learning product and would
go on to be sold to 8 other clients, generating considerable ongoing
income as the cluster are refreshed.
Adapting an algorithm traditionally used to cluster Netflix viewing
behaviors into one used for pharmaceutical prescribing trends makes
for a good headline, the real take away here is that when AI and
machine learning are applied to solve specific issues their impact
can be considerable.
