Machine Learning: The Next Frontier in Commercial Planning

November 14, 2018

Keshia Vaughn looks at how machine learning can transform commercial planning and outlines what teams will need to deploy it effectively.

Chances are, if you are a pharmaceutical executive, you will already have devoted plenty of time and energy to exploring the possibilities of machine learning. Regardless of whether you had been a proponent or a skeptic, the bold claims and non-stop buzz surrounding machine learning will have grabbed your attention and challenged your thinking, demanding that you evaluate its potential and take a position on its suitability for your business.

Machine learning represents a giant leap forward in advanced data analytics. The ability to integrate real-world data sources and build a dynamic picture of markets, patients, providers with such depth, granularity and precision has unlocked a vast goldmine of untapped patterns and insights that could inform (and potentially transform) strategy and drive growth for the business. By creating a dynamic, multi-faceted view of the landscape, machine learning enables the data to speak for itself and reveal the true story, without human bias.

How machine learning can transform commercial planning

In simple terms, machine learning enables teams to look at all the data at once, deploying an algorithm to reveal which are the most important patterns of data. It extends way beyond a linear analysis, offering a dynamic understanding of the market, lending itself to a number of planning activities, such as market sizing, patient segmentation, targeting, provider segmentation, payer segmentation, and messaging, as well as Health Economics and Outcomes Research (HEOR) activities.

Because patients are profiled in a holistic, multi-faceted way, it enables us to see what other diagnoses they have, what healthcare services they are using, what other treatments they are receiving, and any number of other characteristics from the data sources. By building profiles of known patients, we can then apply these to the general population, to find undiagnosed patients. This approach can be particularly effective in underdiagnosed populations.

One example is the NASH (nonalcoholic steatohepatitis) market which includes a largely unaware and asymptomatic patient population. In a recent project, DRG experts deployed machine learning techniques with the goal of expanding the NASH population beyond diagnosed patients and creating meaningful sub-segments. This required an analysis of the progression to NASH to determine key characteristics among potential patients, matching EHR and claims records to build a profile of known patients, and machine learning modeling to flag undiagnosed patients based on key characteristics (i.e., elevated liver function tests, comorbidities).

The analysis found that at least 25% of the NAFLD (nonalcoholic fatty liver disease) patients have progressed to NASH, highlighting the true extent of under-diagnosis within the category. Machine learning enabled the client to build four clinically distinct NASH segments, along with three hybrid segments. It also unearthed one particularly interesting insight, namely, that as many as 10% of NASH patients do not present typical disease characteristics. In cases like this, where the use case is appropriate, machine learning can be an invaluable tool to better understanding and supporting a patient population.

But it doesn’t always have to be a battle between machine learning and traditional methods. There are scenarios when it’s perfectly fine to run the old-school analysis because the business question may not require advanced methodology. In fact, there are cases when machine learning might over-complicated the analysis. It’s important to work with an experienced data-science partner to determine the appropriateness of your case.

Examples of specific business questions machine learning can answer



Illustrative Business Question

How Machine Learning Can Help

Market Sizing

What’s the size of my undiagnosed patient population?

–Build a model from the patients we already know

–Validate and build an optimized model

–Score patient universe to identify patients with high probability of having the disease


Patient Segmentation

How can I create patient segments in a saturated market for a prevalent disease?

–Leverage EHR/claims data to build an algorithm which identifies naturally occurring patient profiles within the wider patient population

–Identify clinical indicators of differentiation between the different profiles

–Apply to the larger disease population to estimate the size of the patient segments within the broader population



How can I target patients with a high probability of having a particular rare disease before they get diagnosed?

–Build a model from the patients we already know

–Validate and build an optimized model

–Score the patient universe to identify patients with high probability of having the disease

–Re-score the universe to alert pre-diagnosis “triggers” within this population

–Identify and message patients’ corresponding healthcare providers


Healthcare Economics and Outcomes Research (HEOR)

How can I build a cost-effective HEOR predictive model that can leverage patient-reported outcomes within EHR data?

–Integrate claims/EHR data to build an exploratory model, which allows us to identify clinical features that differentiate patients who are successful on treatment, or who have high costs

–Validate the model using a hold-out sample

–Apply these bias-free characteristics to optimize the performance the predictive model

–Summarize the results for publication to the broader healthcare community

Provider Segmentation / Education

How can I determine which messages are most effective for engaging HCPs and driving Rx behaviors, without resorting to the biases of physician-completed surveys?

–Integrate EHR, claims and communications data to build an exploratory model that can identify patterns of characteristics associated with physicians’ experiences and courses of action taken in response to different types of brand messaging

–Identify indicators of differentiation between the provider profiles

–Validate and build an optimized model using different types of messaging

–Identify gaps and opportunities for brand messages, and optimize these according to the established characteristics of each segment

–Continue to benchmark model performance and optimize provider profiles and segments

Product Sales Analysis

How can I find out why my launch product is underperforming?

–Build a model that integrates medical and pharmacy claims data

–Determine the factors affecting a sale (e.g. are providers writing it, are insurers approving it, are patients filling their scripts, or are they trying it only once?)

–Apply these characteristics to identify opportunities for removing hurdles to sales and  for targeting providers with brand messaging



What does it take to deploy machine learning? What teams will need

1. Clear objectives, use cases, and output requirements

Machine learning needs human direction. It is critical to know what you want to achieve, what puzzle you are trying to solve, and how you are defining the patient. Therefore, you will need to determine your business question, understand how you might approach the project, define what will constitute success, and think about how you are going to implement your findings.

2. A willingness to embrace the methodology

Machine learning offers a new way of thinking about commercial planning activity and a departure from pharma conventional wisdom. It requires that you approach it with an open mind. You must also be willing to overcome human bias – taking the guesswork out of analytics does not reduce the value of your role, but it does increase the value of your insights.

Machine learning will benefit a variety of use cases, providing the appropriate algorithms are applied in an appropriate manner. But it doesn’t make sense in every scenario: for example, you wouldn’t apply it to finding undiagnosed patients with a common disease, such as diabetes. Machine learning is not a magic bullet. It still requires human expertise to generate insights, apply the analysis, and develop/implement strategy. Successful projects can leverage a mix of traditional analytics and machine learning.

3. Strong partners in data and data science

Machine learning demands that you have access to both great data and great data expertise. If you buy the data yourself, then you risk ending up with modelling sets that are inappropriate for the type of patients you are looking to pull. And if you don’t have access to full, universal data, you won’t be able to find undiagnosed patients.

A good data science partner will guide you through every step of the project and will make recommendations according to your objectives and case uses. They should be able to demonstrate deep knowledge of the healthcare sector as a whole, strong expertise in your specific therapeutic categories and markets, and vast experience of undertaking successful projects using the latest data science technology.

They must also have the professional integrity to base their recommendations on your specific needs, and not on theirs. Sometimes the correct advice will be that your project doesn’t need machine learning at all. Unfortunately, some pharma companies have been burned in the past by working with partners that fall short of these standards.

Stages of machine learning adoption

1. Development

In the initial stage, the data science partner will work with your team to determine the specifications to which they will design and build the initial model.

The data tends to look different for every patient population, so the partner will perform an exploratory analysis to better understand its characteristics related to your target population. Some of these populations will have lots of interactions with the healthcare system, even within rare diseases, but others less so. The partner will make recommendations and will help you determine your specific business question, applying therapeutic area expertise to create the best approach for identifying the patient population. The partner will then build the initial machine learning model.

2. Validation

In this stage, the model will be tested and refined using real-world data. Essentially, it will be tested for its ability to identify patients with a known diagnosis from the previous year. Is the model over- or under-selecting those known patients? Business logic and therapeutic expertise is then applied to help determine which business rules and filters will refine your model and optimize its performance.

3. Implementation

Once the validation stage is complete, and the algorithm has been optimized, your model can go live. You will begin to receive a report, or dashboard, according to the output cycle you requested from the partner (e.g. weekly, monthly, etc). You can start using these insights to shape your commercial strategy, and subsequently implement it via your field sales activities and messaging.

4. Recalibration

It is essential to continually recalibrate the machine-learning model to incorporate shifts in the healthcare landscape. Say, for example, you are undertaking a study within a rare disease market for which initially there were 1,000 known patients. However, in the first year, you have identified and are now treating an additional 250 patients. The experiences and characteristics of these additional patients can be incorporate into the model to further optimize its performance. Essentially, you are continuing to change what that market looks like as you grow that market. If you don’t keep recalibrating the model, you are going to miss opportunities.

Keshia Maughn, MPH, is Director of Data Science at Decision Resources Group.