Top Considerations when Choosing an Artificial Intelligence Project for Healthcare



The idea of employing artificial intelligence (AI) to improve organization performance is capturing the attention of healthcare leaders due to the technology’s ability to analyze and interpret complex unstructured data. However, before an organization jumps into using AI, it should carefully review the pros and cons. Although the technology has broad applicability, it is not appropriate for every endeavor. In part, this is because the industry is still in the early stages of AI adoption, and implementing an initiative can be a substantial undertaking. If organizations aren’t diligent about project selection, they could end up with an expensive and resource- intensive learning opportunity that falls short in terms of meaningful ROI.

How can an organization be confident a project is ready for AI? Here are three important considerations to keep in mind.

  • Look for the biggest opportunities to apply AI. Large-scale projects that are core to an organization’s business tend to yield the most benefits from AI because they are more likely to draw the attention and buy-in of senior management, as well as staff. The initiative should also address a widely-acknowledged pain point, so the time and resource outlay is deemed worth the effort. Although costs can fluctuate depending on the scope of the project, a typical enterprise AI project costs between $1.5 million and $2 million to complete over its full two-year life. Note that costs may decrease as the industry becomes more proficient with the technology, but as of now, it’s best to focus on programs that have a multi-million dollar impact to offset development costs.
  • Check for discernible patterns. At its core, AI builds models that interpret data and discover patterns that convert inputs into outputs. These models tend to produce more accurate and meaningful results when applied to complex datasets that have strong underlying patterns. A simple test of whether a pattern is strong is if a human can identify it. If so, then it is likely an AI algorithm can do the same. A pattern should also persist over time. Otherwise, by the time the AI project is finished, the pattern will have changed, which is a frequent reason why these models sometimes deteriorate after implementation.
  • Be sure there is enough data. To be successful, an AI project must have an extensive amount of data to analyze. For example, if an initiative involves interpreting imaging information, there should be about 1,000 data points per category. Creating a data set of this size is no small task, but it is essential: a lack of sufficient data is a primary reason why AI projects fail. The data also must be clean and labeled so the technology can fully uncover true patterns. Labeling can occur either automatically during normal business operations, such as when medical claims get denied by an insurance company, or manually, by employing a dedicated data labeling team.

The bottom line is organizations must think carefully before applying AI to a project. It is just as important to identify initiatives that are not appropriate for AI as it is to pinpoint the right opportunities. By focusing on a large-scale, core business problem that has solid discernable patterns and an ample data set, organizations can realize project success, achieving significant benefits that have been elusive with more traditional information technologies.

Alex Ermolaev is director of AI for Change Healthcare