In any cross-functional project, it’s important to get executive support from the beginning to enable collaboration and manage expectations. Without the support of executives, it’s difficult to weather unexpected results or lengthened timelines. The project could be doomed to fail before it even begins.

It can be especially hard to establish executive support and get AI projects off the ground because there are few AI project precedents to help determine expected results and predict ROI.

Flagler’s advice is to resist the temptation to offer too many guarantees. Instead, understand your limits and make a contingency plan. You can use the project as an opportunity to learn—from success and failure. Here are Flagler’s tips to pitch an AI project.

Pitching AI projects to executives

• Champion the project as a priority for the entire organization, not just as an IT project, to encourage cross-departmental collaboration.
• Commit to a timeline that gives upper management an out if things aren’t going well but is also enough time to achieve results.
• Recruit a physician to spell out how success will lead to hard ROI.
• Craft your proposal as what you’d like to achieve and what you would learn from failure. For example, Flagler’s pilot led to the discovery of “dirty” data sets and disparate order sets. Regardless of the pilot’s success, the experience helped the hospital improve the quality of their data.

Working with an AI vendor is an investment of both time and money. Even the best products aren’t plug-and-play solutions because every organization is different and will require different levels of support throughout the project. When choosing a vendor, Flagler looked for a true partner to work alongside and one that would help yield the greatest value for both organizations.

Symphony AyasdiAI’s experience in working with other health systems helped them recognize how Flagler’s workflows were different from other systems and where those differences created problems in the code. Where it would have taken weeks for Flagler’s team to solve a simple problem in the code, Symphony AyasdiAI provided an immediate solution by adjusting the code as needed. Relying on Symphony AyasdiAI as a partner helped save time and effort as Flagler prepared to launch the pilot.

Flagler’s team added value to the partnership by confirming clinical results. Symphony AyasdiAI’s solution uses unsupervised learning to recognize patterns in large amounts of data. This is different from supervised learning models where researchers train the algorithm by feeding it specific variables. Although bias is less of a concern in unsupervised models than in supervised models because the algorithm isn’t trained by researchers, there’s still a “black box” problem. Black box problems refer to our inability to understand how the AI makes decisions to create patterns. This is why it was important for Flagler to validate suggested patterns in the data. For example, the solution showed that early use of antibiotics and steroids were effective treatments for pneumonia patients. Even though Flagler’s team couldn’t say how the AI solution made that connection, they could confirm that clinical research supports the early use of these substances to reduce LOS and adverse outcomes.

EHRs are full of valuable data. Flagler realized that with very little work, AI could produce actionable insights from their data. Dr. Michael Sanders, CMIO of Flagler Hospital, thinks of EHRs as the new textbook of medicine—there’s a lot to learn if you can just pull out the right information.

Flagler’s eight-person informatics team wrote 2,500 lines of Structured Query Language (SQL) code to pull the data for the pneumonia pilot from the EHR. Fortunately, the data requirements were well defined and already standard outputs of the EHR, so it was relatively quick and simple to get the data.

Because the AI solution could comb through different kinds of data, Flagler’s team also pulled data from sources outside of the EHR. The other two sources of data were the financial system and the surgical system. The informatics team collected and recorded data for everything that happened to patients from admission to discharge, including standard lab results and vital signs. The team even recorded small details like what meals the patients ate and the exact times they ate each day.

At this point, the solution recognized patterns in Flagler’s data and sorted the patients into treatment groups. Each group represented a typical patient population with similar treatment patterns. The software analyzed the data to create a care path for each group and displayed a dashboard comparing each group by costs, mortality, and care outcomes.

Flagler’s team used the dashboard to identify the optimal treatment group (the group with the best outcomes at the lowest cost) and used the generated care path as a template for a new order set.

The hardest part of the project at Flagler was matching the AI-generated care path to a standard order set. It wasn’t a data or technical problem—the challenge was convincing physicians to change their workflows. Flagler overcame initial hesitation by helping physicians learn from each other, both as project champions and by sharing data comparisons with peers.

Flagler encourages physician adoption and collaboration in all data-driven projects with the help of their Physician IT (PIT) crew. The PIT crew consists of physicians from every department who act as project champions. Whenever IT projects affect workflows, the PIT crew decides if the workflow changes are reasonable, how the change should be implemented, and what the drawbacks are. The PIT crew is then responsible for helping their peers understand the value of a given project.

Knowing that evidence of success is another essential component for physician adoption, Flagler shared patient outcomes and physician adherence to order sets with physicians. The data was especially meaningful to physicians because they could compare themselves to peers, so they could learn from each other. For many physicians, this was the first time they had access to measurable data-driven results.

With physician support, Flagler successfully reduced five conflicting and duplicative orders sets down to one order set.