Strategic Considerations for Adopting Clinical AI

It may be late compared to other industries, but Artificial Intelligence (AI) is now moving rapidly into healthcare, representing a new frontier of innovation in patient care. The rate of investment in healthcare AI is increasing by roughly 40% annually, with an estimated $6.6 billion to be spent on AI by 2021[1]. Besides the logarithmic increase in investment, AI is now mature enough for healthcare that it is demonstrating value. Many uses are now evident, from cybersecurity to managing the supply chain to clinical use cases, and the deployment of AI is already reaping results.

In the clinical setting, there are many opportunities where AI can improve care and promote better outcomes for patients. With access to both medical and external data, clinical AI can identify patients on an avoidable path to deterioration and the interventions that can keep them on the path of wellness. By helping physicians direct their care to modifiable patients, AI can save patients from future harm and save providers from avoidable care costs. When clinicians are chronically overburdened, it’s valuable for both doctors and patients to have AI that can process reams of data and highlight the information doctors need to make the right decisions for their patients.

There are several approaches to consider when planning to implement an AI data-driven solution. These approaches generally fall somewhere on the spectrum between building a solution from scratch and buying a prebuilt solution. Regardless of the provider’s size, successfully building AI tools that bring value is a high-cost, high-effort endeavor. The risks include managing expectations, prolonged runway for first successes, and overall complexity. The build approach also depends on having a mature enterprise data warehouse with well-established processes, broad data sets, and good data stewardship.

On the other hand, buying a solution is appealing because of the perceived rapid speed to value, fewer requirements for analytic resources, and the opportunity to leapfrog the analytics maturation process. However appealing, caution needs to be taken with this approach. With so many new vendors competing in the AI market, few have achieved proven results. Proven results are often substituted with a low-priced “Co-development opportunities,” which sounds very attractive, but more often than not ends in lackluster results. It is important to validate claims by checking references and understanding what is actually a live product vs. what is marketed.

The best approach is typically a hybrid of building and buying. As the analytics team matures in depth and breadth, the demands for analytics will continue to outstrip the capacity to deliver on them. In parallel with that maturation journey, vendor solutions can serve several purposes, including delivering on early wins, the capability to grow with the organization, and providing a core set of value, allowing the analytics team to function at a more strategic level that aligns to broader organizational goals.

The most important strategic consideration, regardless of whether you build, buy, or choose the hybrid approach, is to choose a starting point and to develop a roadmap. It may be tempting to pursue these innovative new solutions simply because they exist. But without proper alignment of strategic initiatives, quality imperatives, and well-defined value metrics, delivering on that value promise becomes increasingly difficult.

Choosing the Right Clinical Solution

There are many clinical use cases where machine learning is applied to create prediction models. Typically, a large data set is analyzed to identify relative risk factors to create a model that is then applied to a new population. These models are promoted and diffused widely to address problems such as LOS, sepsis identification, readmissions, pressure injuries, and more. However, there are several problems with this approach:

  • Over-emphasizing prediction accuracy leaves too many false negatives, and true positives are often already clinically recognized.
  • Models designed to catch false negatives generate excessive alerts, resulting in alarm fatigue and clinical rejection. This creates a tension between producing value and promoting adoption.
  • Risk stratification of only high-risk patients leads to the challenge of how to manage and mitigate those risks. Without intervention insights, the risk is not actionable.
  • Developing protocols, increasing resources, and creating process improvement initiatives subsequently increase costs with little improvement in quality.
  • Last, these models are rarely scalable to a heterogenous population; applying the model to a different region or demographic often results in having to rebuild the model to apply it effectively.

In contrast, AI-enabled clinical prescriptive analytics use sophisticated pattern and machine learning techniques to arrive at some of the same answers by an entirely different method. Patients are risk-stratified based upon thousands of variables through the incorporation of socioeconomic, environmental, and clinical data. These insights provide patient specific interventions that prevent negative outcomes, avoid harm events, and utilize the same or fewer resources.

Consider the case study of a large, multi-region Midwestern health system. By 2020, all-cause readmissions were expected to be in the top decile, with several regions in penalty. After adopting an AI-enabled clinical prescriptive analytics solution in a single region representing approximately 25% of the health system’s total beds, the health system quickly saw results:

  • Averaged 18% reduction in chronic care readmissions (CHF, COPD, pneumonia).
  • $2.9 million avoided in readmissions-associated costs within 10 months of adoption.
  • Over 50% reduction in the variation of care outcomes.

In other hospitals, these same analytics have successfully led to sustained reductions in all cause readmissions by 10-20%, reduction in sepsis episodes by 15%, and sustained reductions in pressure injuries by 40%. Of the many data-driven solutions to emerge for healthcare,  prescriptive analytics is one of the few to have demonstrated value worth the investment.

If you’re considering Clinical AI solutions, download our guide and learn the top 10 questions you should be asking potential providers.

References:

[1] Accenture, “Artificial Intelligence: Healthcare’s New Nervous System.”

Get in Touch

  • 470-427-2900, 3965 Johns Creek Court, Suite 100, Suwanee, GA 30024
  • This field is for validation purposes and should be left unchanged.