There has been a flurry of recent coverage of artificial intelligence (AI) in health careāmost of it flattering. Like the reporting of the success of the BioButtonāused to remotely monitor the vitals of more than 80,000 hospital patients in the past year. There was a recent report on the ability of GPT-4, a large language model by OpenAI to triage 10,000 deidentified emergency room (ER) visits with 89% accuracy. And AI-assisted coding has been reported to improve coding accuracy in the face of the increasing complexity.
But not all the coverage is so positive. For example, The Center for Medicare Advocacy has found that when plans use AI tools, they will issue more denials and more frequent/repeated denials. This has resulted in 32 legislators writing CMS with their concerns and lawsuits filed alleging a lack of monitoring and evaluation of these tools. āAnd, most are leery of the lag between technology advances in the health care space and regulation. A recent study of the use of AI in coding found significant gaps with the highest match rates at 49%. AI-driven mental health therapy (with bots) has problems ranging from rogue ābad adviceā to exacerbation of social isolation.
My take is that AI is likely to play a big role in health and human services going forward. But (like most innovations) there will be as many failures in application of AI as there are successes. As a result, executive teams of community-based provider organizations need to include AI in their evaluation of future technology investmentsāand increase their AI āIQā. The question for these executive teams is how to evaluate the AI-infused technology tools and platforms in a very fluid market space. This question was the focus of the recent session Comparing Healthcare AI Platforms: 7 Must-Haves for Any Behavioral Health Organization. The session was led by Rony Gadiwalla, Chief Information Officer, GRAND Mental Health, Cally Cripps, Vice President of Information Technology & Business Analytics, Aurora Mental Health & Recovery,, Brandon Ward, Psy.D., Chief Innovation Officer & Vice President of Information Systems, Jefferson Center, and Amanda Rankin, Customer Insights Lead at Eleos Health.
The panelists focused on seven key characteristics of ābest practiceā AI-enhanced solutions for behavioral health organizations. Their evaluation framework includes a sector-specific focus; alignment with existing technology and workflows; technology partner support; production of value and performance insights; assurance of data security and HIPAA compliance; ability for customization; and clinician-centered documentation processes.
AI built specifically for behavioral health The speakers emphasized the importance of AI applications built specifically for behavioral health. This includes a framework built on evidence-based behavioral health with specific appropriate language, therapeutic interventions, and documentation models. This specificity is important because the tool must pick up on nuances in patient-provider interactions that a general-purpose AI tool might miss and an AI application built for behavioral health is evaluated and validated by clinicians in its development. Mr. Gadiwalla describes the importance of sector specificity saying, āDo they really, fully understand the context? Can they identify different types of clinical techniquesā¦used during the session?ā
Alignment with existing technology and workflows Integration of any new technology with the existing provider organization platform, reporting structure, and workflows is important for any technology acquisitionāand AI is no exception. This alignment allows unified organizational performance reporting. It decreases support time. And it saves staff time and promotes ease of use. Mr. Gadiwalla talked about this in terms of sustaining product adoption as follows, “Alignments with existing tech is critical. Right? It can make or break a project. There are plenty of great ideas, especially with start-ups, but a very small amount of those can actually be operationalized…Maybe a year after deployment..can you keep this thing running?”
Support from the technology partner When selecting technology, executive teams need to assess what support they needāand what support the technology vendor can (and does) supply including customization. The speakers discussed how this includes development of realistic timelines and expectations about implementation and training requirements, as well as mapping out the responsibility for ongoing support and maintenance. Mr. Gadiwalla gave context to this support saying, “They brought their passion to the table. They engaged our clinicians to help them understand how Eleos would work in our environment… And I think that really made the difference between a customized product and something that was picked up off the shelf.”
Value and performance insights A big value of any technology investment is the derivative dataāand the information and insights provided to optimize organizational performance. In evaluating AI-enabled technology platforms, executive teams should evaluate the ability to access data in a format that provides those insights. Ms. Rankin described the importance of performance data saying, āYour AI platform should help youā¦providing objective data on adoption, time savings and staff satisfaction. Otherwise, how are you going to know if itās working?ā.
Data security and HIPAA compliance A key area for executive team evaluation of any AI-enabled technology is the vendorās plan for data security and HIPAA compliance. The vendor should be able to produce security certifications (e.g., HIPAA, SOC 2, HITRUST) and describe their security protocolsāaudit controls, privacy, security controls, and integrated third-party applications. In addition, any contract with the technology vendor should specify the process and responsibilities for managing any data breach or disaster. The contract should also include data retention protocols specific to the provider organization. Mr. Gadiwalla mentioned that the rapidity of AI tool development by startups may leave important security considerations as an afterthought. He talked about trust, saying, āOur providers need to trust that the information is being used responsiblyā¦the folks that we serve have to knowā¦that weāre not using it for reasons that are not communicated.ā
Clinician-centered AI documentation A final key characteristic of best practice AI-enabled technology is ease of clinician useāand executive teams should evaluate the technology from this perspective. Some of the features in AI-enabled technology that executives should evaluate include the clinical language, the ability of the solution to evolve based on user feedback, and how the technology interacts with clinicians and how/when their approval of content is required. āIt actually allows for folksā¦to be really, really present in the session,ā says Dr. Ward.
While provider organization adoption of AI-enable