AI Assistant Helps Prepare Patients for Cardiac Cath

A voice-based artificial intelligence (AI) assistant named Sofiya has been helping prepare patients for cardiac catheterization at a high-volume center in New York City, freeing up nurses to focus more on direct patient care.

In a two-phase pilot study, the preprocedural call completion rate approached 90%. More than one-third of calls—which included the provision of instructions for the patient and collection of clinical information—were handled entirely by Sofiya without the need for escalation to a nurse, researchers led by Annapoorna Kini, MD (Mount Sinai Hospital, New York, NY), report in a paper published recently online in npj Digital Medicine.

They estimate that using Sofiya instead of a nurse saved an average of 11 minutes per call, or the equivalent of 37.3 12-hour nursing shifts over the course of a year.

Amid shortages of healthcare workers, AI has been seen as a way to shift nurses—and in the future, physicians as well—away from routine tasks like making preprocedural calls and back to the bedside working with patients, which could help improve healthcare workers’ mental health and wellbeing, Kini told TCTMD.

“A lot of this mundane work can be taken care of by this . . . conversational agent,” she said, indicating that an AI assistant could eventually also take some time-consuming tasks—like completing charts and dictation notes—off of physicians’ to-do lists in the future. “[In] all these places AI can . . . help them to facilitate and finish their work faster, so they’re not taking their work to their home and they can recover faster and spend more time with their family.”

For now, the current study shows that a “conversational AI agent can be safely used,” Kini said. She added that Sofiya was trained “with a lot of guardrails because this was first of its kind,” noting that if patients ask about something that is not included in the AI knowledge base, they receive a call from a nurse.

That’s a key safeguard, Kini explained. Patients “should not feel that AI is completely taking over their care as long as they know there’s a human in the loop and this AI agent is assisting the healthcare workers, . . . so they are able to focus on more clinical work.”

Commenting for TCTMD, Thomas Maddox, MD (WashU Medicine, St. Louis, MO), said the study is “a good illustration” of how AI can augment current processes rather than replace the human element of care.

He highlighted the fact that the majority of calls completed during both phases of the study required some involvement from the nursing staff.

“It wouldn’t be correct to say, ‘Oh, this is replacing what the nurse is doing,’ but it certainly is augmenting what he or she is doing,” Maddox said, adding that the estimates of time saved by using Sofiya need to be validated. “But I think it’s directionally correct that they did generate a fair amount of time savings per patient interaction, and over time that could translate into some real efficiencies.”

Vetting Sofiya

This study was a prospective evaluation of the initial use of Sofiya for conducting preprocedural calls at a Mount Sinai cath lab performing more than 16,000 procedures each year. Sofiya is based on a large language model and deployed on an agentic AI framework. The day before a procedure, Sofiya calls patients to provide instructions on diet and logistics around getting to the hospital and to collect required clinical information, including allergies, medications, and more. If patients ask a question that Sofiya can’t handle, a nurse follows up. After the call, an AI-generated report is created and, after verification by a nurse, added to the patient’s electronic health record.

There were two 90-day periods involved in the rollout. The first phase was a stabilization period involving collaboration among nurses, physicians, and AI developers, while the second was led by nurses.

Overall, 1,431 patients (mean age 67.1 years; 28% women) received 1,606 calls from Sofiya—806 in phase 1 and 800 in phase 2. The overall rate of successfully completed calls, defined as those reaching the end of the script with the patient answering all clinical questions, was 86.4% in the first phase and 87.9% in the second.

Overall, 36.6% and 42.6% of calls in phase 1 and phase 2, respectively, were completed without any human intervention, 37.6% and 33.5% required a protocol-driven callback from a nurse, and 12.2% and 11.8% were appropriately converted to a call with a human. Rates of refusal to speak with the AI assistant were 4.3% and 2.8%, respectively.

Calls most commonly couldn’t be completed due to patients not answering or hanging up. An AI system reliability issue resulted in noncompletion in 6.0% of cases in phase 1 and 3.0% of cases in phase 2. Most AI issues were technical in nature, with only six instances of AI hallucinations (ie, Sofiya made things up or provided incorrect information) leading to an incomplete call.

The researchers assessed patient satisfaction using postprocedural surveys among those who completed a preprocedural call with Sofiya. Overall satisfaction rates were high at 94.7% and 98.1% in the first and second phases, respectively.

Some people “love to have a companion to chat with,” Kini said, noting that Sofiya is able to have longer conversations with patients compared with the time nurses can dedicate to the task.

Maddox said clinics at his center use AI for note-taking and the vast majority of patients give their permission to do so and are comfortable with the process. There are some patients who are not comfortable with having AI involved, he said, “but like with any new technology, as they encounter it more and more in their lives—for healthcare or shopping or customer service or wherever they’re seeing it—I suspect the acceptability will go up over time.”

Patients should not feel that AI is completely taking over their care as long as they know there’s a human in the loop. Annapoorna Kini, MD

Asked about the challenges of implementing an AI assistant like Sofiya, Kini pointed to the process of getting numerous committees within the health system, including those focused on legal and ethical implications, to sign off on it.

She and her colleagues mention other potential obstacles in their paper, including “lack of knowledge about AI, concerns regarding its reliability and accuracy, issues related to data privacy and security, worries about the impact on patient relationships, and fears of job loss due to automation. AI education and training, technical assistance, active involvement of clinicians in AI systems development, institutional support, and regulatory oversight can help overcome some of the barriers and facilitate the implementation of AI in healthcare.”

One issue that remains to be clarified is the cost of using AI for this purpose, Kini said, adding that once Sofiya can be streamlined into the electronic medical record, costs should come down.

Looking to the Future

In the future, it’s likely the AI assistant can expand into providing information to patients before discharge, making and changing appointments, conducting follow-up calls after a procedure, guiding patients to cardiac rehabilitation, and more, Kini indicated. Sofiya may also be used for patients undergoing other procedures, such as those for peripheral or structural disease, in the cath lab.

“AI is here,” Kini said. “Everybody should embrace it and try to start using it where you think that it is appropriate.”

She underscored the need for careful implementation. “We are using it cautiously, and we have already shown that patients like it and physicians like it, so we are just gradually expanding the horizons,” Kini added.

As the role of AI in clinical practice continues to expand, there will be a need to minimize hallucinations, ensure that the privacy of patient data is protected, and adjust workflows to allow humans and AI to work together seamlessly, said Maddox, who is a member of the US Food and Drug Administration’s digital health advisory committee.

The healthcare system is “famously slow” at adopting new technologies, but AI is “definitely ready for prime time,” he said, adding that it is becoming a mature technology for the types of applications studied by Kini et al, like callbacks and routine coordination of care.

But Maddox envisions a much broader use of AI in clinical care in the future.

“What I’m really excited about is getting it ready for actual involvement in care delivery. So are we using it to better educate our patients, better prepare them for their clinic visits? On the clinician side, [can it] help us further augment our diagnostic reasoning or our treatment plans?” he said. “And that’ll take an added level of rigor and safety and quality and all those things you would always need to have in place when you’re talking about direct patient care decisions.

“But we’re seeing a lot of progress in that domain, and a few years after we fully inject this into the administrative side of healthcare, we’ll be ready to start putting it into patient care directly, and I think that’ll be great,” Maddox continued. “I’ve been using it where appropriate already, and it’s just super impressive what it can do.”