Value-Based Care Success for Pharma: Requires Machine Learning for Patient Adherence

sgruber-200 (1)By Sarianne Gruber
Twitter: @subtleimpact

How pharma rethinks its marketing will be on par with how healthcare will embrace value-based care. Succeeding in value-based care may indeed be the impetus for pharma to embrace new “technologies” to help patients better adhere to medication regimens.  Speaker Kevin Troyanas, SVP Marketing Analytics from Saatchi & Saatchi Wellness, at the IIEX Health Conference gave an excellent overview of why machine learning and predictive analytics has become the technology of choice for healthcare marketers to tackle patient adherence.

Patient Adherence: Essential for successful outcomes
“Healthcare marketers are now addressing what it means to incorporate this value-based ‘thing’,” imparted Troyanas.  However, this “thing” is tied to a vested financial interest for both health providers and hospitals to effectively improve outcomes like reduced readmissions.  Even after discharge from the hospital or receiving a prescription post a physician visit, a patient may need better care coordination in order to follow a designated care plan. Healthcare stakeholders have full awareness of the hurdles to adapt and achieve the benefits of a bundled-payment program.  Capped health services during a clinical care episode and preventing penalties, has led to instituting new workflows and patient care plans. Patient adherence to follow-up appointments, educational resources, and medication adherence is very important to avoid readmissions. Pharma marketing companies are also motivated to help patients succeed in their care and wellness. And perhaps, the biggest change has been to relook at the patient data for answers.

Embracing the World of Analytics is Crucial
Predictive models are all around us and you may not even realize their influence in your day to day life.  “Maybe you have received a message that there is the possibility of a fraudulent claim on your credit card,” Troyanas shared as an example.  And in jest, he continued that there isn’t a little elf in your computer viewing you all your transactions, but an algorithm that has been trained on past fraudulent claims generating a probability estimation for potential future fraudulent behaviors. You can even “chase a face” on Facebook.  Image recognition models have become extremely popular.  How does one know that the probability vector of pixels is a particular image? The answer is machine learning, which is a type of artificial intelligence that allows computers to learn how to model data without being pre-programmed.  These algorithms have the ability to predict a correct image with a very high degree of accuracy.  Data science is becoming more and more common in the healthcare space.   Thanks to the increase and the accessibility of data from EHRs and HIEs, we can even predict hospital readmissions before they occur.

Machine Learning: A Democratized Technology
“We are really seeing an explosion in machine learning techniques in the industry,” pointed out Troyanas.  The democratizing of this technology and its tools is being powered by the open source environment, including the programming languages Python and R that are absolutely free.  The same phenomenon is going on with algorithms.  New machine learning algorithms are being developed in academia such as neural networks, extreme gradient boosting, tree algorithms and random forests. We are also seeing a high demand for Data Science studies being taught in academic environments as well as online courses.

What is interesting is the how the data can differentiate the type of machine learning algorithm to use.  There are two big categories: supervised learning and unsupervised learning.  We use supervised learning methods on data when we know what happened for at least a certain group of individuals. For example, if a dataset contains patients that “lapse” in their therapy, we can build a predictive model using supervised learning methods like regression and neural networks.  The other case is unsupervised learning when you don’t necessary know what you are looking for, but instead, you let the machine learning algorithms (such as hierarchical cluster, association maps, and self-organizing maps) find patterns inherent in the data itself.

Predicting Risk of Prescription Lapse
Starting at the time a patient gets the first script, he or she can either persist to the second script or lapse.  For the majority of brands, persistency curves show that half of the patients’ lapse by the 6th month.  Usually, the first to second transition has the largest drop-off.   How do you know which patients have the highest risk to lapse?  The answer is uncovered with machine learning.   An optimal modeling approach would be a Markov chain, where the probability of next event only depends on the previous event and the not sequence of prior events.  Training lots of algorithms require longitudinal data containing prior long-term behavior of patient adherence. Those patient characteristics or descriptors affecting the target variable are used to convert the patient type into a probability of lapsing.  The model of choice is the one that best predicts lapse. Then you can look at the “whole universe of probabilities” to find patients who are most likely to lapse. Essentially, you want to develop risk assessments of your patients and be able to cluster them into segments. Unequivocally, Troyanas advised that the highest risk patient segment is where to initiate an early spend and effort for behavioral change.

Learn more about the Insight and Innovation Exchange Health (IIex Health) Conference that was held April 4, 2017, in Philadelphia.