By Beth Friedman, Sr. Partner, FINN Partners
LinkedIn: Beth Friedman
X: @FINNPartners
Co-Host of FINN Voices – #FINNvoices
The same day White House officials announced a second round of industry commitments for the “safe, secure, and trustworthy use and purchase of AI in healthcare,” a group of more than 200 healthcare professionals met to discuss the identical topic (and a few more) during the HIMSS AI in Healthcare Forum. Held in San Diego, the event featured speakers from high-profile organizations including Microsoft, Cleveland Clinic, and the American Medical Association.
Dozens of industry speakers explored every aspect of AI and machine learning (ML) during the two-day event. Some sessions offered hopeful predictions. Others shared proven examples of AI and ML deployment, technology advancement, and clinician adoption. These real-world uses of AI and ML caught my attention, and inspired attendees’ confidence.
Each case study session contributes to the library of lessons we’ve collectively learned so far in healthcare’s next great conquest—delivering on the promise of AI to improve health outcomes.
Six Proven Tips for Successful Deployment of AI in Healthcare
Talking about AI’s potential in healthcare is one thing. Learning new lessons from scalable deployments is another. Here are six valuable takeaways from hospitals, health systems, and vendors reporting progress with AI and ML.
Get started. Even taking small steps with AI and ML is essential for healthcare stakeholders in 2024, cautioned Chris Larkin, CTO, Concord Technologies. Larkin recommends that organizations “tackle a small component of a single business problem that’s worth solving” and “show investors the result of your work in weeks, not months or years.” Many speakers throughout the event reiterated Larkin’s plea to test the AI waters.
Use a problem-first approach not a technology-first approach. Sunil Dadlani, Executive VP, CISO, and Chief Information and Digital Officer, Atlantic Health System, emphasized the need to accurately define specific business problems to solve before evaluating any technology solution. Dadlani shared 13 AI use case categories the organization reviewed before narrowing their initial focus to operational efficiency in the clinical documentation process. Atlantic Health System’s first deployment addresses clinician burnout by improving documentation efficiency through ambient voice and Nuance’s Dragon Ambient Express (DAX).
Focus on operational efficiency. Addressing the administrative burdens of clinical documentation and other operational processes was a common starting point for many of the speakers. Healthcare’s mid-revenue cycle was identified as a proven area to realize value through AI and ML.
Harjinder Sandhu, CTO, Microsoft Health and Life Sciences, also recommended starting with administrative use cases. Luis Ahumada, Director, Health Data Science and Analytics, Johns Hopkins All Children’s Hospital, reiterated the prioritization of administrative and operational streams for AI use. The organization’s biggest impact with AI and ML has been experienced in clinical documentation.
Lars Maaloe, CTO, Corti, goes a step beyond simply relieving administrative burdens in documentation. The company works with thousands of clinicians to improve physician-patient conversations. This includes alerting clinicians to documentation gaps in real time, suggesting more empathetic responses, and fine-tuning the AI tool for each physician over time. To accomplish these goals, systems must keep up with the live clinician and patient dialogue. “We watch, listen, and measure physician feedback so we all get smarter together.”
When selecting which administrative problem to address first, Sandhu reminded attendees that each organization is different. Every use case carries a different risk profile, varying rationale, and unique considerations. Efforts should deliver high value at low risk to start. He suggests that healthcare organizations ask the following questions related to administrative use cases:
- Is it work the risk?
- Does the cost to implement make sense?
- How do you enable the use case and make it scalable?
One attendee at the event, Matt Zubiller, CEO, e4health, reiterated the importance of ensuring consistency across teams and ramping productivity through automation with specific focus on the messy mid-revenue cycle. “Bots create consistency of data decisions in the revenue cycle. By digitally analyzing the data, AI quickly identifies problem areas for expert human teams to manage by exception.”
Address data quality early and often. Dadlani also mentioned the importance of quality data. Selection of a properly built and medically curated library is an important step to ensure AI and ML outputs are valid and valuable to the end user. According to Severence Maclaughin, CEO and Founder, DeLorean AI, there are variances in healthcare’s large language models (LLMs) and outcomes shift as data changes over time. Organizations must avoid training LLMs with bad data.
Furthermore, LLMs differ based on location, region, and patient population. According to Forrest Pascal, Principal, AI/ML Model Governance, Kaiser Permanente, “Data may never be fully clean because humans are involved, but we can do our best to make sure the data is as unbiased as possible and we serve as good data stewards.”
Ann Cappellari, MD, CMIO, SSM Health, reminded attendees in a later session that, “garbage in, garbage out” also applies to AI. She encouraged the industry to “take care with automation, and balance how much you want the system to learn versus starting from a very rigid pattern.”
Finally, Ahumada reiterated the need to address data quality before implementing AI or ML. “The data we’ve been collecting for more than 30 years isn’t the right quality to create the new ML models we want to use.” Ahumada encouraged a renewed focus on specific data problems and closing data collection gaps.
Engage end users from the start. According to several nurses and many physicians attending the event, the industry must avoid adding keyboard clicks to clinical or administrative workflows. “End users must be part of the AI conversation at the very beginning,” said Sophia Henry, Clinical Consultant, Beckman Coulter Diagnostics. Their knowledge can help developers build better products and systems.
Rachini Moosavi, Chief Analytics Officer, UNC Health, emphasized that “the hardest part of AI implementation is integrating new systems into end user workflows.” Physicians are especially sensitive to the addition of new technology. Eve Cunningham, Chief of Virtual Care and Digital Health, Providence, suggests the following tips related to end user adoption of AI or ML:
- Think ahead to how AI or ML solutions fit into the end user workflow.
- Accurately estimate the demand from the end users. What is needed to train and change processes to adopt the tool?
- Is the right technology and human infrastructure in place?
In conclusion, Christopher Sharp, CMIO, Stanford Health Care, mentioned the need for aligned incentives with clinicians, departments, and the entire organization. “If incentives don’t align, adoption won’t advance.”
Deliver value quickly. Cunningham reminded attendees that perpetual AI pilots aren’t sustainable. While there is no shortage of great ideas, efforts must be prioritized to achieve value and then continually measure return on investment as the solution is implemented across the organization.
Looking Ahead to HIMSS AI in Healthcare Forum ’24
As Harvey Castro said during his opening keynote session, “There is too much on the line for us to get AI wrong in healthcare.” Castro added that there are no experts. We are all learning as we go.
Events like HIMSS AI in Healthcare Forum continue to educate all of us on the promises and practicalities of AI in healthcare. And as today’s guidance evolves into tomorrow’s regulatory rule, there likely won’t be any facet in the healthcare industry that won’t be touched by AI.