Editor’s note: This is part of a series of articles sharing best practices from companies on the road to become model-driven. Some articles will include information about their use of Domino.
In North America, pet owners spend more than $62 billion (as of September 2019) caring for their cats and dogs. Without medical insurance for their pets, owners often face financial stress and difficult decisions when their beloved companions are injured or become ill. I’m proud to work for a company that’s applying a model-driven approach to help pet owners afford the best veterinary care possible: Trupanion. In this blog, I’ll describe one of the ways we’re applying models to impact the pet health insurance experience, and share insights into how’ve we built and managed the data science team that’s delivered them.
But first, a bit of context
Even when owners have medical insurance for their pets, they often must pay for services out of pocket and wait weeks to find out if a test or procedure is covered. There’s much uncertainty in the process and often significant financial strain when vet bills reach hundreds or thousands of dollars. (In fact, while an average claim is around $300, it’s not uncommon for some vet invoices to total $10,000 or more).
Trupanion wants to remove this uncertainty by not only providing comprehensive coverage but also by using algorithmic models to simplify the claims process. Our proprietary and patented software is currently installed in approximately 4,000 of veterinary practices in the US. Through our software, vet bills can be processed and paid directly at the point of service in less than seven seconds. Pet owners only pay their portion of the bill, which is typically 10% of covered services less their chosen deductible. They no longer have to submit claims or wait for approvals and reimbursement. It’s a game-changer for both our customers and their pets, helping vets to recommend and pet owners to select the best course of treatment for their pet regardless of the cost.
One challenge we faced in developing our software to automate the claims process was accounting for the many different ways that veterinary practices record services. There aren’t standardized medical codes in veterinary medicine as there are in human medicine. And because our pets can’t tell us what hurts when they’re ill, medical diagnosis can often be vague, such as “not feeling well today.”
There’s a lot of interpretation and expertise that goes into assessing a claim, and we had to mimic this human decision-making. So we created a series of independent models, each designed to analyze a specific item or issue with a combined predictive accuracy of 99%. Some models use simple text search to identify non-covered items like pet toys. Others use deep learning frameworks to uncover connections among words across one or more visits. We currently have 15 models working in parallel to adjudicate a claim. They each have their own code repositories, version control, and run on their own hardware in Amazon Web Services (AWS). Results from each of these independent models are then fed into a final decision model that correlates these predictions and decides whether to pay a claim and what deductibles to apply.
In building our Direct Pay system, we also focused on the people-side of the equation, particularly in two areas:
- Effectively leveraging both expert data scientists and citizen data scientists
- Addressing automation fears head-on
Leveraging both expert and citizen data scientists
We have a small number of analytics experts—with only a handful of data scientists, BI analysts, and other data experts. To achieve the level of automation needed and grow the use of models across all areas of our business, we need to tap into each group’s expertise and provide learning opportunities for all. I’ve focused on several initiatives to this end:
- Building a cohesive community: Our Data Science and BI teams sit in close physical proximity, leading to an organic sharing of ideas and knowledge. Our data scientists are responsible for developing and deploying models, such as those automating the claims resolution process, detecting fraud or predicting customer churn. But if a business analyst has an idea for a model, they can confirm feasibility of that idea using the autoML tool DataRobot. We encourage business analysts and other “citizen data scientists” to own these modeling projects. The BI team is constantly working with the data behind the metrics most important to the company. As such, they are in a great position to generate ideas on where predictive models can improve processes or even change the way we look at something. So far, BI has worked on email classification, net promoter score (NPS) prediction and lead scoring models. During development, our expert data scientists actively coach our citizen data scientists, advising on feature engineering and providing best practices throughout the coding and model evaluation process. Given that claims automation takes up about 70% of data scientist’s time, this community-led approach enables greater experimentation than would otherwise be possible.
- Ensuring consistency and collaboration: I encourage business analysts to use the same tools that our data science teams use; DataRobot primarily for prototyping models and Domino for coding and managing models through production. There are several benefits: easy access, knowledge transfer, and ease of transition.
- Emphasizing creativity: Both our Data Science and BI teams split their efforts using the 70/20/10 resource allocation philosophy. They spend approximately 70% of their time on work that we know is valuable, 20% on projects that have clear potential, and 10% on speculative projects. In fact, claims automation was originally a speculative project, and now it is core to our business. Innovation is in our DNA. In order to remain industry leaders, we must be willing to test ideas and try things that no one else has done before. This approach helps ensure the long term health of our company and our teams as we continue to grow.
Addressing automation fears head-on
We have hundreds of people in our organization with immense medical knowledge whose sole job is to resolve claims. As a company, we need to communicate that the system isn’t replacing their jobs, but rather augmenting and freeing them up to focus on complex issues where they can make the most impact. Additionally, our experts are essential to continually improving our software, in essence training the software with every decision they make.
In addition, we’ve looked to help adjusters resolve more difficult claims with greater speed and efficiency. For example, we’ve built a model-driven decisioning tool to help guide claims adjusters through particularly difficult and time-consuming decisions. Adjusters using this tool have realized a 20% increase in efficiency so far.
We’re off to a great start, and expanding
Already, we’ve been able to automate more than 75,000 claims at 4,000 vet practices in the US, paying $6.1 million in benefits. We’re improving our customers’ overall experience while reducing costs.
With more than 25,000 vet practices in the US alone, we’ve still got a way to go. However, by focusing on the people part of the equation and putting in place processes to help tap into everyone’s knowledge, we’re well on our way to achieving our goals.