AI can drive changes in health care but won’t replace human decision-makers

Although the term was coined in 1956, never before has artificial intelligence (AI) had such a moment. Interest and application of AI in many forms from machine learning to generative (e.g., ChatGPT) is booming across industries, including health care.

A recent survey by Morgan Stanley Research showed that 94% of health care companies use artificial intelligence/machine learning in some capacity and that the health care industry’s average estimated budget allocation towards these technologies is projected to grow from 5.7% in 2022 to 10.5% in 2024.

We spoke to several local professionals in health care and academia to find out how AI can assist medical providers and what considerations are important to keep in mind when using it.

VisualDx

Rochester is home to VisualDx, a clinical decision support system used by more than 2,300 hospitals, clinics, and medical schools around the world. Founded in 1999 as Logical Images Inc., its mission is to improve medical decisions through augmented thinking and timely visualization.

Using AI machine learning capabilities, the company provides equitable medical images to medical providers to help them with diagnosis. This is something especially helpful for providers who are not specialists in a particular field, like dermatology.

“There’s a great global need for dermatology and pattern recognition because most skin disorders are not treated by dermatologists,” explained Art Papier, M.D., a dermatologist and co-founder and CEO of VisualDx, “They are treated by primary care or analyzed by the emergency doctor.”

In that situation, Papier says a non-dermatologist may not be sure what the rash they are seeing is and if it’s something serious. That’s where VisualDx can help by using AI and machine learning to give another tool to non-specialists so they can have more expertise as they work.

“When you show a rash to a non-expert, like a generalist care physician or nurse practitioner who doesn’t have expertise in dermatology, they’re only accurate 36% of the time in describing the skin exam,” Papier said. “We’ve improved that accuracy to 68% when using our technology. Our technology doesn’t replace the doctor, but it improves their skills dramatically.”

VisualDx also offers tools for consumers, like Aysa, a skin rash app and symptom checker for iPhone and Android where users can snap a photo, answer a few questions, and receive guidance on what to do next.

Since its inception, VisualDx has been committed to improving the diagnosis of skin conditions in patients of color. A 2020 study by the Journal of the American Academy of Dermatology showed VisualDx was the leading provider of diverse dermatological skin images from a review of top online and printed texts worldwide; 28.5% of all images in VisualDx are of dark skin.

The company has 80 full-time employees made up of software engineers, imaging specialists, medical librarians and other support staff, as well as 50 physicians on its editorial board.

“We’re a Rochester story because if you ask what this town does, well, we do health care really well and we also do computer science really well,” Papier said. “We have incredible talent from the University of Rochester and R.I.T. and a lot of great competencies in AI and machine learning.”

URMC

At the University of Rochester, Dr. Conrad Gleber, M.D., M.B.A., is a hospitalist and assistant professor of medicine who has a strong interest in AI. He’s an active participant in the digital health innovation incubator at URMC called the UR Health Lab.

At the UR Health Lab, medical providers and researchers collaborate with data analysts, computer scientists, engineers, business professionals, and more to create tools and systems that use advanced machine learning models and other technologies.

Among the aims of the UR Health Lab is to transform health care and drive change via technology and strategy and to ensure a commitment to ethics and mindfulness throughout the innovation process.

Gleber’s interest in artificial intelligence started in business school around 2012 during discussions about data. Currently, he is involved in research projects centered around AI and classification, ambient documentation, and trustworthiness (how to make AI models that are aware of their own bias and resistant to bad actors).

Ambient documentation — a type of ambient clinical intelligence — is like having a digital scribe. It uses highly advanced voice-enabled AI to automatically document conversations between a provider and patient during visits.

Gleber is not using ambient documentation or any other type of AI in clinical practice, but he is interested in how AI could potentially lessen the administrative burden of being a physician and give physicians more time to see patients.

“Being a doctor is a very different job than it was twenty-five years ago — the burnout rate is growing,” said Gleber, who explained physicians entering the field now have more administrative tasks than ever before and often less interest in doing them.

Gleber believes that as opposed to operational-type AI tools, clinical implementations of AI are not going to be mainstream anytime soon and, if and when it is used, it will be another instrument the physician has and not a decision-maker.

“Clinical tools will be in synergy with the doctor,” Gleber said.

RIT

Dr. Linwei Wang, Ph.D., a professor in RIT’s Golisano College of Computing and Information Sciences and director of RIT’s Personalized Healthcare Technology research center, agrees that while machine learning has the potential to make medicine, it will never replace physicians and their thought processes.

“For clinicians, AI is one more piece of information,” Wang said. “It doesn’t replace their decision-making process.”

Wang also directs the Computational Biomedical Lab at RIT where part of her research is focused on how machine learning can improve patient care, especially in cardiology.

She has written about the risk of a one-size-fits-all perspective when it comes to deep learning in medicine due, in part, to potential bias for subpopulations underrepresented in the training data.

Wang says that while AI has very good potential to improve health care across a very large spectrum, it’s important to be aware of hidden biases in data.

Caurie Putnam is a Rochester-area freelance writer.