Predictive analytics works in the real world without us understanding exactly how. We turn on our cell phones and expect them to work without understanding the fundamentals. We push the gas pedal and expect the car to go. Even something as simple as a pencil may not be common knowledge as far as production is concerned.
If you don’t understand how predictive analytics works, you’ll never move forward to new models and paradigms. However. Dr. Phillipp Diesinger and colleagues from pharmaceutical company Boehringer Ingelheim outlines how we are now complementing system modeling with something more statistical, machine learning, to overcome difficulties with both systems in this talk from ODSC Europe 2018.
[Related article: Announcing the First ODSC Europe 2020 Virtual Conference Speakers]
Two Cultures: Machine Learning and System Modeling
According to Diesinger, 98% of statisticians are using regression models, with only 2% regularly using decision trees or neural networks. In system modeling, first principle models or differential equations rule most models.
On the machine learning side, challenges such as limited system insights, irrelevant theory, and questions about whether conclusions are about model mechanisms and not system mechanisms are huge concerns for improving predictive outcomes.
System modeling, on the other hand, could have a vast number of approaches, too many to operationalize into any business-scalable solution. They’re analytically challenging and resource-intensive without the benefit of universal applicability for business impact.
Businesses need these analytics, but without improvements in our understanding of their real-world application, neither may answer the question fully.
For example, a business wants to build a population estimation tool to predict the number of potential attendees at a conference. In machine learning approaches, you would collect your data (conference location, venue size, ticket price, marketing, etc.), perform some feature engineering, and then train and evaluate your model.
In the system modeling approach, you use identity first approaches. For example, conference badges are numbered consecutively. The chances of running into someone with a four-digit number is relatively low for a conference with 1050 participants but would be common for 10,000. You can use a mean-field probabilistic solution.
The Approach in Real Life
Let’s take a look at two approaches the pharmaceutical company is taking in real life to create predictive models that account for two unique obstacles in healthcare/pharma.
Prescriptions and Marketing with Dr. Gabriell Mate
A real-life example is one that Boehringer Ingelheim is deeply familiar with. As a pharmaceutical company, it relies on sales-rep oriented marketing, much more so than other fields do. The industry is heavily regulated, and marketing initiatives are highly resource-intensive.
Optimizing this marketing initiative helps keep prices lower and ease pressure on doctors. It’s tough to use traditional approaches because data is anonymized and unavailable, and conventional economic metrics fail in this particular field.
Prescriptions tend to follow a system of collective behavior. There’s a well-defined order that moves prescriptions from a disordered to an ordered state, so can we use that to better predict marketing efforts for such an unusual field?
This phenomenon is called the “phase transition.” The company wants to use this concept to apply to the spread of opinion across a network to influence that opinion within reason.
- Ising Model: Describes moving from a disorganized state to organized with two states.
- Potts Model: Generalization of this same phase transition model for more than two states.
In our search, we understand that the doctor’s opinion depends on a few different states: current opinion of the disease, opinion of their peers, and as an outside influence, marketing. We need to understand the opinion of the doctor’s network and marketing actions.
Unlike medical data, we can find plenty of data about our doctor’s network: alumni associations, workplaces, research collaborations, and others. Combined with generalized prescription data, we can get a much better idea of the state of things.
The model changes the state of doctors based on random probability (the Markov Model). The model produces a matrix that predicts how doctors influence each other for prescription levels and the potential influence of marketing on this network.
There are some significant insights from this process:
- doctors influence each other
- those opinions matter more than marketing
- there is a threshold below which marketing doesn’t matter
- the threshold changes from doctor to doctor
- the effects of marketing saturate fast
Pharmaceutical companies can use this to simulate marketing effects for a more significant impact despite the challenging obstacles presented by healthcare data.
Demand Elasticity with Dr. Berenice Pila-Diez
Predictive models must also account for healthcare’s unique take on supply and demand. Pharmaceuticals aren’t direct to consumer, and supply isn’t quite as limiting of a factor. For demand, there are quite a few differences in things like competitors, popularity, and effectiveness.
The company uses perturbation theory to get a better idea of this demand elasticity. With this modeling, you use a system with a known variation and add correction terms or perturbations to solve for the potential variance.
We can start to understand how demand evolves organically over time. As you add simple additions like price, you begin to understand more about how the complex demand for something that doesn’t follow traditional supply/demand concepts.
System Modeling for Complex Problems
The company has used system modeling to handle challenging predictions for the pharmaceutical company, including predicting marketing and supply and demand. Check out their talk for a specific walkthrough for both of these real-life examples.