5 practical steps to successful clinical AI deployment

Organizations treading carefully incorporating artificial intelligence into clinical workflows do so with good reason. Healthcare faces impediments to AI implementation, namely cybersecurity and data governance, while AI tools still struggle to accurately interpret clinical risk. 

With that in mind, a recent report from the Commonwealth Fund highlighted how to make AI useful for clinicians. The report pays particular attention to two things: How to separate practical and pie-in-the-sky ideas, and how to earn clinicians’ trust amid pushes for AI scalability. 

AI deployment is easier said than done

Successful AI deployment requires testing and refining tools in real-world settings. That tends to pose five challenges for hospitals and health systems. 

• If the decision about where to invest comes down to clinical vs. financial benefits, and there’s no formal process for comparing the two, then ROI will likely prevail – and create a rift between financial interests and clinical needs.  

• Electronic health records are comprehensive yet often incomplete and frequently biased. Organizations may need to augment EHR data with socioeconomic and community-level data to accurately and equitably model unmet clinical needs. 

• Staff who see how well an AI tool solves one problem may want to use the tool to address a second issue without realizing such purpose-built models will need to be retrained on new data sets. 

• Patients don’t always trust AI; along with concerns about clinical misdiagnosis, they worry that the output of AI-aided documentation may limit access to care, increase out-of-pocket costs, and inadvertently leak their personal data. 

• AI deployment remains an n=1 phenomenon since clinical environments have different patient populations, staffing levels, and financial models. As Duke Health’s Dr. Mark Sendak put it, “AI hasn’t had its penicillin moment” that results in widespread impacts on clinical outcomes across care settings. 

Key AI lessons from successful health systems

A recent survey found 75% of health systems are using at least one AI application. The Commonwealth Fund report noted the most prevalent clinical tools emphasize patient engagement and documentation. For example, nearly everyone is using ambient notetaking, while 93% are automating in-basket management and 80% are augmenting remote monitoring with AI. 

The health systems that have moved beyond these use cases have five things in common.  

Balance local and enterprise decision-making. Organizations such as Duke Health and the Mayo Clinic vet AI tools at the enterprise level but let individual facilities decide whether to adopt those tools. In addition, Vanderbilt Health has created an interdisciplinary team that helps assess whether AI models work as intended and/or may require further oversight to mitigate risk. 

Link to public and local data sets. PCCI, a spinoff of Dallas-based Parkland Health and Hospital System, gathered clinical data from 100-plus health systems and combined it with data sets outlining 26 social drivers of health. This has helped the organization do things like map disease burden and document correlations between SDoH and clinical outcomes – and do so more accurately than using clinical data alone. 

To build trust, show effectiveness. Experienced clinicians are used to assessing risk based on what they know and may be skeptical of a model’s output. Duke Health leaders demonstrated the effectiveness of decision support tools that helped predict increased risk of HIV diagnosis by showing how the AI model identified patients that traditional risk-assessment criteria would have missed. 

Roll out new models slowly and watch performance closely. Duke tests certain applications in “shadow trials;” that way, they can run in the background without influencing clinical decision-making. Vanderbilt monitors its 300-plus live AI models in a real-time dashboard and flags unexpected outcomes such as signs of bias or model drift, both of which limit a tool’s effectiveness over time. 

Engage staff early in the design process. AI tools tend to be more dynamic than software products that are built once and infrequently updated. When end users understand AI and feel confident using it, broader adoption is more likely. These feedback loops also make it easier to redesign clinical workflows with AI in mind, rather than simply adding another technology layer onto an already-flawed workflow. 

Brian Eastwood is a Boston-based writer with more than 10 years of experience covering healthcare IT and healthcare delivery. He also writes about enterprise IT, consumer technology, and corporate leadership.