Dustin Tart can measure improvement in treatment of one of the deadliest conditions at Duke University Hospital by a single time marker: Before Sepsis Watch and After Sepsis Watch.

Before Sepsis Watch, an artificial intelligence tool created at Duke Institute for Health Innovation and implemented at Duke University Health System to track and treat the life-threatening response to infection, Tart saw firsthand as a Nursing Program Manager how tiny changes signaling a turn toward sepsis were missed, and treatment was delayed.

"We didn't really have a standardization around sepsis care," Tart said. "It was kind of sporadic."

Sepsis is sometimes colloquially called "blood poisoning," and occurs when an existing infection triggers a chain reaction that can rapidly lead to organ failure and death. It is the No. 1 cause of death in hospitals, according to the Centers for Disease Control and Prevention.

The problem is that many initial symptoms, such as fever and elevated heart rate, mimic other illnesses that even the most experienced clinicians might miss.

"When you're talking about sepsis, it's subtle," said Armando Bedoya, a Duke Assistant Professor of Medicine in Pulmonary, Allergy and Critical Care Medicine who was instrumental in the development of Sepsis Watch, along with Dr. Cara O'Brien.

This is why the idea of developing a deep learning model to provide Duke with an early warning system seemed like an intriguing solution.

Sepsis Watch was trained by inputting data from 42,000 patient encounters with 32 million data points. The artificial intelligence deep learning system helps synthesize all that information and compares it with 86 real-time variables from current patients every five minutes, flagging who might be susceptible to sepsis.

After Sepsis Watch was introduced in 2018, deaths attributed to sepsis dropped 27%, and use of the AI tool has expanded from the emergency department to the entire Duke University Hospital.

"A lot of people develop AI models, but not many are integrating them into clinical practice to improve clinical outcomes," said Suresh Balu, Associate Dean for Innovation and Partnership at Duke School of Medicine and Director for DIHI. "That is a huge differentiator for us at Duke."

Sepsis Watch was one of Duke's earliest successful forays into artificial intelligence use, but the growth and accessibility in the past two years of generative AI – through software such as ChatGPT – has ushered in an explosive use of AI at Duke.

AI is used by a researcher who thinks it can help find a cure for a rare pediatric cancer, by the Nasher Museum of Art to curate an exhibition, and by a Duke faculty member who hopes an AI version of himself will allow for 24/7 office hours.

And that's just skimming the surface of use across Duke, which hosted its first AI summit in April about the wide range of research, teaching and operations activities engaging AI at Duke.

"It's moving quickly toward being a pretty basic skill set," said Evan Levine, Senior Director for IT Services and Support in Duke's Office of Information Technology. "We're going to want to make sure people know how to use it correctly, to know where the pitfalls are."

When the Nasher Museum of Art had a hole in its exhibition schedule last year, Chief Curator Marshall Price joked they should ask ChatGPT to plan something to fill it. But the curatorial staff didn't laugh off the suggestion. What if they really did? they wondered.

That's how the Nasher's "Act as if you are a curator: an AI-generated exhibition" came to be. The exhibition opened in September 2023 and ran through February 2024. Creating it wasn't as simple as asking the chatbot to conjure an exhibition from the 14,000 objects in the Nasher's collection. That first query yielded an answer that included works of art "not even in our collection, let alone a real artwork," said Julia McHugh, a Curator and Director of Academic Initiatives.

With help from Mark Olson, Associate Professor of the Practice of Art, Art History & Visual Studies at Duke, a custom interface was built that included the Nasher's full public-facing catalogue. Even then, AI made some interesting choices under the theme it selected for itself, "Dreams of Tomorrow: Utopian and Dystopian Visions."

AI created ethereal and dreamlike landscape portraits filled with depictions of lush vegetation.

"It's like you are bringing your camera into a time machine," Forte said, "and as soon as you shoot it, you see something that makes you say, 'Oh!' It's magic."

The computer screen in Pranam Chatterjee's lab in the Pratt School of Engineering is filled with neon-green squiggles, thick sections of fusilli-pasta-like springs and narrow lines that curl in every direction. It's a digital model of a protein in the cell of a rare pediatric cancer called alveolar rhabdomyosarcoma, or ARMS.

Chatterjee, an Assistant Professor of Biomedical Engineering, describes it by the very technical term, "spaghetti monster." He and his lab are trying to design something that will attach to the irregular shape, and they've used generative AI to help.

"No one can touch this protein," Chatterjee said. "And that's why this cancer is not cured. But as you can see, our binder is centered right inside that protein. So, it is able to hit it and kind of embed itself within that protein."

Hitting the protein with a binder created with the help of a generative AI algorithm means that protein can be destroyed. If the protein is destroyed, the cell will die.

Chatterjee is fairly certain that AI is going to help cure this cancer.

Chatterjee's lab has trained its custom-built generative AI by feeding it billions of points of data specifically related to protein in cells, so that when they write complex algorithms asking AI ambitious questions, they're confident in the answers.

"That is literally the hardest protein on the planet to bind to and to destroy," Chatterjee said. "If we can get to that protein, I think we can do anything – and this would prove to us that our technology makes a difference."

Chatterjee and his lab collaborators are using AI to examine some of the rarest and most difficult diseases: ARMS and Ewing Sarcoma, two rare pediatric cancers; Huntington's Disease, a fatal neurodegenerative disease; and Alexander Disease, a rare neurological disorder. They develop models for binding to the proteins, and then test them in the lab – a rare one-stop shop.

"Those are our favorite targets because they're just hard," Chatterjee said. "Generative AI is probably one of the coolest ways to design drugs for them."

Tony O'Driscoll did some quick math recently to illustrate a problem. About 420 students took the Duke Adjunct Professor's classes in both the Fuqua School of Business and Pratt School of Engineering in Spring 2024. It was very common for most of them to ask for at least 15 minutes of his time outside of class.

He offered 90 minutes of office hours per week over the 14-week semester.

That meant 21 available office hours – but a demand for 105 hours. It didn't add up.

"Could I create a digital copy, or clone of myself?" he asked.

After about a year of working on it, he plans to launch ProfTonyAI this fall.

"If this technology allows a student to get access to meaningful and useful content at their 'moment of need,' then I think it adds value," O'Driscoll said. "That teachable moment could be at 3 in the morning when I'm asleep. My digital version doesn't sleep."

In every presentation Steve Toback gives on artificial intelligence – and he gives them often because it's one of his passions – he includes a quote that has been attributed to various people across the internet: "AI won't take your job. Someone using AI will."

The underlying message from Toback, a Duke Media Architect and Senior Producer for Academic Media Production in the Office of Information Technology (OIT), is that everyone needs to learn how to use AI responsibly and ethically.

Using Microsoft Copilot is one way. All staff, faculty and students have access to Copilot (formerly Bing Chat) at no charge by using their NetID to log into the platform. Using a Duke login ensures that any data entered remains secure and is not publicly shared.

Microsoft Copilot might be most useful with basic tasks: Summarizing long emails or web pages, finding just the right formula for a difficult Excel task, creating an illustration for a Zoom presentation.

But receiving the best support requires giving Copilot (and other generative AI applications like ChatGPT) clear instructions or questions, Toback said.

"The better-formed your question is, the better response you're going to get," Toback said. "Prompt engineering is everything."

Prompt engineering means giving the chatbot parameters and expectations for the information you're seeking, and includes starting by giving it a role with the prompt, "Act as if you are..."

But even if the best prompt is provided, AI answers are only as good as your own knowledge of a subject – because generative AI answers are often not correct. It's still not an expert on everything but will answer as if it is.

"Broad-use generative tools are kind of good at a lot of things," said Levine, OIT's Senior Director for IT Services and Support.

Advancements in AI are evolving quickly with potential for transformative change.

"Duke is doing what Duke does best, which is innovate in all sorts of different areas and spaces," Levine said. "There's a lot of really cool stuff happening in the right ways, and I think we'll learn from that and come together for bigger and better things."

How are you using AI at work? Let us know by writing working@duke.edu.

Follow Working@Duke on X (Twitter), Facebook and Instagram and subscribe onYouTube.