A whole body of human bones containing data about the molecular structure and particle background of the genomic DNA.
getty
Many of us close to the AI industry have heard a lot about healthcare applications. But there are many ways to talk about this. You can use an anecdote. You can focus on a specific setting (rural hospital, free clinic, family clinic) or one specific deployment (e.g. radiology or drug discovery). However, in some cases, using a wide-angle lens can be beneficial.
Let’s start here. What is “continuing care”?
It is explained as follows Science Insights article:
“Continuum of care is a framework that describes the full range of health care services that may be needed over time, from preventive testing to treatment, rehabilitation, and long-term support.”
We also received the following information from Copilot, courtesy of Bing’s AI Overview:
“Continuum of care is a coordinated system of services that ensures that individuals receive consistent and seamless support over time and across different levels of medical and social services.”
Do those two look like you? The first configures the CoC as a scope of service, and the second configures it as a system of services. For me, the first one is a little more specific.
Either way, if we think of the continuum of care as the full range of services across the patient lifecycle, we can think about all the ways AI can be applied.
Input from panel
At the Imagination in Action event in April, a panel discussed these applications. Whoop’s Emily Capodilupo interviewed many experts. Whoop took the device to the FDA, so there’s some relevance there. (Disclaimer: The April IIA event is an annual conference that I facilitate.)
The group was thinking about how to make these AI applications work properly and how they fit into the future of healthcare.
Terrifying consequences of AI
Capodilupo’s first question to the panel was: “Why is AI keeping me up at night?”
“Trying to build solutions for the next day keeps you up at night,” said Eric Rosenthal, program director at the MGB NeuroAI Center. “What worries me is that the models currently in place in healthcare often don’t work.”
He discussed several use cases in imaging where the results may not match expectations.
“We’re not building a robust system that you would call a full-stack trusted AI system. We’re collecting data from a basic mode in a way that we believe the model is reliable, and we’re writing papers and manuscripts that are getting a lot of attention. People are on stage talking about how good it is, but they don’t work.”
Eli Lilly’s senior vice president Gokul Radhakrishnan had a more optimistic outlook on the FDA process.
“They’re pretty open,” he said of the FDA. “They’re asking: How do we enhance the AI to be more deterministic in the validation process? They don’t want a black box, they want a glass box.”
Sharing his own concerns, he explained some of the ways in which deeper AI implementations can function as “more than just a chatbot.”
“I think the area of highest risk is where real-world evidence reaches doctors at the time they prescribe a treatment,” he says. “How do we actually get the right information to the doctor at the right time for treatment?”
Sanford Health CIO Brad Reimer had key concerns related to his role.
“From a bad actor perspective, the potential impact of AI on cybersecurity is definitely a big concern,” he said.
Reimer also discussed “constraint points,” which consider how new AI tools build on previous changes.
“So many processes are built around electronic health records that were forced into place 20 years ago, and that ecosystem hasn’t fundamentally changed,” he said. “It’s hard for us to step out of the box and rethink it.”
“What scares me the most is the underlying data that AI relies on to derive the outputs we want and make clinical decisions based on,” said panelist Kinetic CEO Sufyan Chaudhry, describing an inconsistent process in which too much prompt-based coding can lead to, in his words, a kind of “cognitive dissonance.”
“All of a sudden, we started accumulating cognitive debt as well. So we’ve now moved organically away from prompt-based coding for engineers to more specification-driven development, where we design what engineers can and can’t do,” he said. “Right now, we are a very technical company, and I fear that most of healthcare is not.”
Chaudhry painted a further picture, imagining government intervention.
“The more objective the data the AI relies on, the better the output will be,” he said. “If you rely on subjective data for inference-based output, I think the consequences are going to be dire if you don’t fix it. That’s why protocols are important and why governments need to intervene.”
Reimer said governance is often immature.
“With so much duplication happening so quickly, we have to think about where governance proactively handles AI rationalization,” he said.
Learn more about government
Mr. Capodilupo asked the panelists how government involvement would work.
“Governments don’t have the talent needed to understand AI at the pace it’s growing,” Chaudhry said. “So I really hope that a lot more public-private partnerships come out of that.”
Mr. Reimer had insight from the practitioner’s side.
“You can’t just walk in and let them manage things because they don’t understand,” he said. “At the end of the day, we are still physicians, and medical practices are responsible for patient safety and have a responsibility to ‘do no harm.’ This unknown gap does not allow us to impose regulations to minimize risk.”
enter the examination room
Rosenthal brought up something that many of us have experienced. Applications like Scribe AI, often implemented over the past year or so, allow doctors to talk about our health without having to look at a screen or tap a keyboard.
“Some of the AI is making us more human and connected,” Rosenthal said of the change.
He also mentioned simulations of comatose patients and high-risk scenarios, where families opt-in and often benefit from the insights at hand.
Is there a shortage of doctors?
Chaudhry cited the shortage of doctors as an issue.
Radhakrishnan agreed to use a helper like Scribe, but the visualization was quite different.
“I think a recent study from Deloitte showed a 15% increase in hourly turnover to two to three hours per physician,” he said. “We’re reducing the amount of documentation we do each day. That’s huge.”
He called the new system “ambient” and compared the previous system to “a little box in your chest” that “kind of freaked people out,” a Darth Vader-like device.
“But it’s pretty ambient now,” he concludes.
Will clinician skills atrophy?
Capodilupo then asked the group an almost verbatim question. The general idea is that as more processes are delegated to AI, people will become accustomed to playing a more peripheral role.
“We now have well-trained human doctors and well-trained AI. This combination is very powerful because they really complement each other,” she explained.
In response, Rosenthal told an anecdote about using AI for analysis that matched his own findings, which further observations eventually revealed, earning him some praise.
“We can be led down the wrong path,” he said of the impact AI will have on human clinicians, contrasting it with reading. “But if you take a ‘trust but verify’ approach, this is the new medical education. It’s just streaming.”
Reimer said incorporating change management into training is essential.
Radhakrishnan expressed bullish sentiment on AI that will substantially shape medical outcomes.
“I think we will definitely have medical-grade AI someday,” he says. “It’s just a matter of time, and we need the government and all facilities to help us get there. We’re going to have more medicines. We’re going to have more therapeutic areas with the existing infrastructure that we have today. We need to rely on this, leverage this, leverage AI. “That’s not an option anymore.”
Finally, Capodilupo asked the group where red flags appear in this process of bringing more advanced AI online.
Mr. Radhakrishnan cited a lack of trust that often seems to impede further progress.
“The pharmaceutical industry itself has big challenges,” he says. “You can’t really interact well with patients if you feel like you have to hold a 10-foot pole to even touch them.”
He called for new protocols to advance medical AI.
“How do we actually work with regulators to create that protocol?” he asked. “How can we create trustworthy, medical-grade AI? And how can we actually do it?”
Rosenthal pointed to the need for companies to uphold regulatory standards, such as using NIST-compliant architectures and avoiding liability for data leaks to “countries of concern.”
“It commits us to this kind of collaborative platform,” he said. “And I think one of the things we can do as a group is commit to a platform that we can build together as a data commons or a marketplace.”
“We need a trusted knowledge base,” Chaudhry added, stipulating a codified system and knowledge base that supports robust AI in this area.
I found all of this to be beneficial. Many of these ideas could be leveraged if we are to overcome the obstacles facing medical AI. stay tuned.
