The new pre-procedure coordination assistant at New York City-based Mount Sinai Hospital’s cardiac catheterization lab may not be human, but she knows how to communicate like one.
“She has a soft, calm voice. She’s available 24/7, 365 days a year. She doesn’t call out sick and she has the patience of an angel,” Annapoorna Kini, MD, director of the cath lab and of the hospital’s interventional structural heart disease program, told Becker’s.
Sofiya is also an agentic AI, employed to call patients prior to stenting procedures to answer logistical questions and provide preprocedural instructions.
Dr. Kini spoke to Becker’s about the impetus behind Sofiya and how the technology has been embraced by care teams and patients.
Editor’s note: Responses have been lightly edited for clarity and length.
Question: What inspired you and your team to develop an AI-powered solution like Sofiya for pre-procedure coordination in such a high-stakes clinical environment?
Dr. Annapoorna Kini: During the pandemic, we faced a shortage of nurses and physicians. Projections from various government agencies indicate that by 2030, this shortage will worsen.
Nurses handle both clinical work and a lot of nonclinical clerical work, especially since EMR documentation became standard. This includes repetitive tasks that nurses did not go into the field to do. We need nurses at the bedside, doing clinical work.
In areas like cardiac catheterization, a text message doesn’t cut it. We need to ask specific clinical questions, that’s where artificial intelligence comes in. AI is perfectly suited for those repetitive but critical tasks — like calling patients with instructions, directions and answering common questions.
When we decided to use agentic AI, we didn’t just take something off the shelf. We partnered with a company and fully customized it. I was involved, along with a team of healthcare professionals working closely with computer engineers to build exactly what we needed. While many companies are popping up with AI solutions, employing this in a real clinical setting is different. A healthcare professional must be an integral part of building the solution.
We even chose the voice, it had to be soft, calm and empathetic. A lot of work went into it. That’s probably why it succeeded: We invested the time and effort. We gave the engineers exact questions and answers, and a supervisory AI orchestrated the conversation. Sofiya is more advanced than a standard bot; it’s capable of handling dynamic conversations and circling back when patients change topics.
For example, during a call, a patient might ask clinical questions then suddenly ask, “Which floor am I supposed to go to?” The AI needs to pivot smoothly and return to the checklist.
We’ve reviewed over 4,000 minutes of calls — with some calls lasting more than 30 minutes — and the patient satisfaction rate is over 95%. The kind of patience to wait, repeat and stay engaged is rare; no human would have the patience for that.
Q: What have been your biggest challenges in integrating conversational AI into pre-procedure care and how did your team address them?
AK: When we were starting, we were skeptical ourselves. The concerns were understandable — privacy, ethics and potential job displacement are major topics in healthcare.
Physicians were also worried. They questioned whether an AI could handle pre-procedure calls effectively. What if the patient didn’t understand? Would they still show up? The instructions had to be accurate and reliable, and that raised legitimate concerns.
There were also implementation challenges. In a large healthcare system, you have to ensure data privacy and security. But we had already tackled these issues years ago when we implemented an app to manage heart attack patients long before AI was mainstream. We were familiar with the process of integrating new tech in a large clinical setting.
As for challenges, they mostly relate to accents, pronunciation and speech speed. Some people talk fast, some talk slow. We’re still refining those things and it will keep improving.
As for adoption: First, healthcare staff were skeptical, but they got used to it.
Second, we didn’t know how patients would respond. To our surprise, they liked it. Only about 1% to 2% of patients didn’t want to speak to an AI and hung up.
Keep in mind, our patients are not just younger tech-savvy people. Our oldest user was 94. Most of our heart patients are 60 and older. This technology clearly resonated across age groups and worked well overall.
Q: Sofiya has reportedly saved over 200 nursing hours in just five months. How has this shift in workload changed the day-to-day experience of your nursing staff?
AK: One great thing about our AI agent, Sofiya, is that she can handle 15 to 16 calls at the same time. A human can only make one call at a time.
Burnout has also been an issue, especially with repetitive tasks. When nurses are stuck doing documentation and unwanted phone calls, they start to think, “Why did I become a nurse for this? I want to work with patients.”
Once we handed this task to Sofiya, the nurses loved it. With Sofiya, a day full of calls takes just two to three hours. Nurses now have time to support us in other areas where we need their clinical expertise.
After each call, Sofiya generates a summary. Nurses review it and call the patient back if follow-up is needed, but that follow-up call is shorter and more focused.
Q: How do you ensure that the quality and accuracy of Sofiya’s communications match the clinical standards of Mount Sinai?
AK: In every healthcare system, there are multiple oversight groups that review and approve these tools. You can’t just build something one day and roll it out the next, that’s not how healthcare works.
In the first 90 days, we did over 800 calls. Every single call was reviewed by three of us to make sure nothing was missed and patients weren’t getting upset. That’s also when we ran a survey, and the feedback was overwhelmingly positive.
One thing we were careful to include: At the start of every call, Sofiya says, “I’m a virtual agent,” and at the end, she reminds patients they can always reach a real person. That reassurance matters, especially early on. Maybe in five or six years everyone will be used to AI agents, but right now people need to feel that clinicians are still available if needed.
Leave a Reply