Why It Matters
The generative artificial intelligence (AI) products that have evolved since the public debut of ChatGPT might be a game-changer for health care quality teams, but organizations need to carefully consider the costs and benefits of their use.
In the not-too-distant future, a workday for a quality improvement professional might look something like:
Morning: You prepare for a meeting with your chief quality officer to discuss your team’s ongoing PDSAs. Your secure generative artificial intelligence (AI) “assistant” searches your files and inbox for relevant information and creates a 300-word summary. You realize that you have not updated your run charts with data from the past two weeks, so you provide the data to the AI tool and ask it to generate an updated run chart. You now have all the information you need to make a succinct presentation.
Midday: Five new clinical guidelines for high-volume services in your hospital have been published. (The guideline development and enactment process — from evidence review to approval — used to take at least a year, but it now only takes a few weeks because guideline developers use AI to support background research and data analysis.) You meet with clinical leads to devise a strategy to incorporate these new guidelines into practice. The AI tool you consult recommends small tests of change and a measurement strategy to assess progress. Everyone at the meeting contributes fully to the discussion. AI transcribes and summarizes the meeting and distributes the minutes.
Afternoon: You need to analyze the results of the quality management coaching sessions your team has been leading with clinical teams over the past year. You use an AI tool to review the transcripts of the approximately 500 sessions. The tool uses pre-existing frameworks to identify gaps in the coaching sessions and themes from the participant feedback. You then prompt it to write a 500-word summary of the analysis and build the outline of a 10-minute presentation for tomorrow’s meeting with your quality team.
Learn more about Artificial Intelligence at the IHI Forum
We believe this AI-enabled scenario represents the likely future of QI work. Over the course of a 90-day Innovation project, an Institute for Healthcare Improvement (IHI) research team concluded that all the technology to perform these activities already exists. The generative AI products that have evolved since the public debut of ChatGPT in November 2022 might be a game-changer for health care quality teams, but organizations need to carefully consider the costs and benefits of their use. We arrived at three key concepts during the innovation cycle that we share below.
What Is Generative AI?
Generative AI tools differ from previous technology in their ability to generate new content. In other words, the tools take existing content and can answer questions (in paragraphs), as well as analyze data and provide recommendations and conclusions. Large language models such as ChatGPT, Google’s Bard, Claude, and Microsoft Bing, have attracted millions of users in recent months while also generating controversy. To train large language models, developers used vast amounts of existing content — mainly from the Internet — to process and generate new content. In response to a query, they can provide responses and point users to existing and potentially useful information.
Generative AI Can Support Most QI Modalities, Including Teaching
While our research suggested most practitioners are only in the early phases of using AI tools for QI work, these technologies will likely dramatically improve over time and change how we do QI in the coming years. We can already use AI tools to create most QI materials. Our research found that large language models (like ChatGPT) can help savvy users build run and control charts, identify change ideas, craft cause-and-effect diagrams, and draft driver diagrams.
AI tools can also help teach QI concepts. For example, it can offer explanations of complex ideas tailored to a specific audience. AI tools can draft lesson plans, course outlines, icebreakers, and much more. Some quality specialists have started to use AI tools to visualize data and tackle basic QI questions, such as generating a preliminary set of change ideas or producing a plain-language explanation of a QI concept.
AI Is Error-Prone and Risky
Using AI comes with legal and ethical risks. There has, for example, been an onslaught of litigation questioning whether it is legal to use copyrighted works without authorization to train large language models. AI could also compromise your patients’ or organization’s privacy. It is not permissible to share protected health information with a publicly accessible model like ChatGPT because any data input becomes the property of the company that owns the tool. Similarly, if you share, for example, strategically important organizational data with the tool, you might find the information reused in unexpected and undesirable ways. Prior to using any AI-powered tool to support your work, it is important to consult your IT policies and data privacy or legal teams.
Accuracy is also a significant concern with current AI tools. ChatGPT-3.5, for example, the freely available version, only has data through January of 2022; it is already out of date. AI models have also been known to “hallucinate” — to make up plausible-sounding information in response to a prompt. In addition, language-based models do not perform consistently on mathematical tasks (e.g., ChatGPT got worse over time at identifying prime numbers).
Additionally, AI models pose equity concerns, given their training on large data sets that are likely to reflect the biases of society at large. While major tools like ChatGPT put safeguards in place — like Reinforcement Learning with Human Feedback (RLHF) that use human supervisors to train the models and guide them away from unsuitable results — they can still fall short and produce content that could be offensive or biased.
It is unclear what impact these AI tools will have on the quality and safety workforce. Thus far, the most successful users have deployed the tools to complement and extend their work. The tools are not mature enough to fully replace any specific workflow in a typical QI manager’s day.
Recommendations
We recommend that leaders consider the following actions to respond to the growing interest — and clear risks — of using AI tools:
- Convene multidisciplinary groups (including representatives with legal, IT, clinical, quality, and research expertise) to develop specific governance and acceptable use policies for new large language model-style AI tools. For example, many organizations will only allow staff to use “enterprise” versions of the tools (designed for secure business or government use) for work-related tasks.
- Train staff on your organization’s acceptable use policies.
- Train staff on how to use generative AI via “prompt engineering.” The outputs of these tools are only as good as the inputs, and “prompts” refers to the text questions or instructions entered into the system to get the desired results. Numerous resources are coming online to help users ask the right questions.
- Collect staff feedback on AI. Ask how they are already using it. Inquire about why they are enthusiastic about it or why they have qualms.
- Help staff integrate AI into their current work with an awareness of the possible ripple effects, including how it might enhance productivity, change day-to-day tasks, and shift employees' suite of responsibilities. In introducing AI into daily work, leaders and staff will need to be nimble, for the available technologies are changing rapidly, and new tools are being introduced at a rapid pace.
Marina Renton, MPhil, is an IHI Research Associate, Innovation and Design. Pierre M. Barker, MD, MBChB, is IHI's Chief Scientific Officer, Institute for Healthcare Improvement. Gareth S. Kantor, MD, is a clinical consultant, Insight Actuaries & Consultants. Jeffrey Rakover, MPP, is an IHI Director, Innovation and Design. To learn more, attend session A24 “When AI Meets QI” at the IHI Forum (December 10–13, 2023).
Artificial Intelligence in Health Care: Peter Lee on Empathy, Empowerment, and Equity
Everyday AI Podcast: How AI Turns Clinical Trials into Medical Knowledge