HIT Think

4 keys to success with AI and machine learning

While artificial intelligence and machine learning have revolutionized other industries, they are still newer concepts in healthcare.

AI is all around our daily lives now. Consider some of the personalized experiences we get from algorithms working behind the scenes in other areas; for example, “People You May Know,” and “You Might Also Like.”

Many clinicians would love to have this type of experience when they open their electronic health record (EHR), tailored to meet their specialty-specific workflows and documentation styles. The algorithms behind AI and machine learning will eventually transform healthcare, especially through enhanced population health management, healthcare access and quality.

HDM-061918-AI (1).png

What will it take to fulfill the promise of AI and machine learning in healthcare? There are four things that will help usher in a new era.

Expand use within clinical decision support
Clinical decision support is based on evidence and research. It’s often hard-coded into systems to deliver more standardized, best-practice care for patients. Today, clinical decision support embedded in the EHR will remind clinicians to record height and weight to calculate BMI, or to perform medication reconciliation at each office visit.

But with machine learning, clinical decision support can do so much more. We can transform systems laden with meaningless alerts to intelligent workflows and best practices driven by relevant patient history.

Best practices should change as new evidence becomes available, and it can take a long time for new findings to affect patients. Sadly, some estimates suggest that it can take as long as 20 years before research findings become part of widely accepted clinical practice. Algorithms can change that by enabling more timely, practical and precise clinical decision support.

Look for proactive applications within population health management
Machine learning can enable clinical decision support based on multi-system analysis to understand which patients are at highest risk of a negative outcome, or to optimize treatment in real-time. For example, organizations can be more proactive in identifying and protecting against infections.

Algorithms can parse available historical and current information to inform clinicians which patients are at risk for specific outcomes or deliver personalized treatment plans for patients with chronic conditions. For instance, guidelines recommend 30 minutes of aerobic exercise five days a week to patients, but the plan for a patient who has multiple comorbidities—such as individual with angina and Type 2 diabetes in addition to COPD—may need to account for markers for safe exercise tolerance.

Build trust with clinicians
Adoption and acceptance of these technologies is slow. Clinicians are trained to validate recommendations with evidence. It is important to expose the “how” and “why” behind AI and machine learning algorithms to the clinicians who will use them. Clinical decision support algorithms should be subject to the same rigor we have for academic research.

EHR vendors can build trust by practicing good habits that will build trust and speed adoption. And while we should be protective of intellectual property and the algorithms themselves, we have an obligation to be transparent about our development processes, data sources and algorithm performance.

For example, we have found that 80 percent of prediabetic patients progress to diabetes. Once prediabetes advances to diabetes, it can cause serious health complications. But early intervention can help prevent or delay the onset of diabetes and improve outcomes for patients.

Any clinician will ask questions about how we reached this conclusion, especially because this is an unexpected finding much higher than the Centers for Disease Control and Prevention (CDC). They are much more likely to accept this estimation of risk if they know our:

  • Development process. We collaborated closely with experts from the American Diabetes Association (ADA) and created rules with the CDC criteria that define prediabetes.
  • Data sources. We analyzed about 50 million de-identified patient records and combined it with other types of data, including claims data, consumer information and environmental data.
  • Algorithm performance. We applied this rules engine to our EHR data and followed these patients for four years.

With understanding of how we reached this conclusion, clinicians will be more willing to agree that the risk for prediabetic patients is indeed high for progressing to diabetes and should have intervention. They are more likely to agree with clinical decision support that encourages clinicians to enroll these patients into diabetes prevention programs.

Make insights more helpful to the caregiver
Cool algorithms are great, but they must be something the end user wants and can use. Insights should be embedded in the technology they’re already using to be successful. Organizations should also be prepared to act on findings. Unless there are intervention programs in place to help address predictions, the information is not impactful.

AI and machine learning should enable:

  • Precise problem selection, enabling clinicians to pick the best possible diagnosis code for better clinical decisions (and more accurate billing) downstream.
  • Intuitive workflows, giving a more humane, medically sensible documentation process to reduce clinician burnout.
  • Real-time risk stratification, helping clinicians develop better care plans.

Ultimately, insights from data is only worthwhile if we deliver them, in a consumable way, to the point of care.

The next few years will require a lot of work to bring about successful AI and machine-learning transformation. We’re using our data lake of 50 million de-identified patient records as “training data” to further develop and hone our algorithms. We’re working closely with select clients who want to use insights from their data to improve population health management and access to care for their patients.

When we can demonstrate high performance rates, we can deploy these algorithms on a bigger scale. Only then will we begin to understand the full capabilities of AI and machine learning.

For reprint and licensing requests for this article, click here.