Unlocking the Power of AI

IN RADIOLOGY, a field with a shortage of physicians but an explosion of medical images waiting to be read, AI helps lighten the workload. Currently, algorithms spot minute bone fractures in X-rays, suspicious spots on mammograms, lung nodules and brain bleeds on CT scans. Those images get pushed to the top of the pile and potential lesions are flagged for evaluation by keenly trained human eyes. Medical residents still learn time-tested ways of assessing images, but they also learn to use AI tools that will make their work lives easier.

A similar burden is being lifted in pathology, which often entails conducting meticulous and time-eating counts of cells through a microscope. But is tackling that tally the best use of highly trained brain power? Now, AI can pick up some of these mind-numbing tasks, freeing pathologists to focus their energy and expertise on complex analysis that improves patient care.

There are also vast troves of data created in the research of Alzheimer's disease: PET scans and MRIs, patient histories and biological samples. With the help of specially trained AI algorithms, experts have a powerful assistant in their search for early indicators of the disease and potential drug targets.

When it comes to cancer, AI aids researchers and clinicians in conducting deep analysis of genomic data. Through this sifting process, they can find early cancer markers, which might better explain how a disease unfolds or — ideally — offer a new target for therapy. When combined with clinical information, researchers can tailor treatments to each patient and even forecast the probability of cancer returning. AI is aiding discoveries in breast, kidney and prostate cancers, as well as myeloma and melanoma, to name a few.

In pediatric cancers, which, like Trent's, are inevitably rare, large clinical trials that are standard with other diseases are a challenge to put together because there simply is a lack of patients. Even so, AI is helping researchers search for answers in animal models.

Artificial intelligence is also slowly filtering into exam rooms. Ambient listening records conversations between a doctor and a patient, but it goes beyond producing a transcript. When built into the software for an electronic health record, it can handle some essential documentation, extract critical details from a conversation, and write concise notes. The technology not only eases a physician from the burden of charting, which many often complete at home, but frees them up to look their patient in the eye.

The key pivot point in AI for IU School of Medicine is the Department of Biostatistics and Health Data Science, which supports more than 200 research grants around IU School of Medicine, many with AI components. "We mine large amounts of data — genomics data, imaging data, clinical data," said Kun Huang, PhD, the department's chair, "and we generate hypotheses. We make predictions."

Newer AI models are graphically connecting the dots between data points, said Jing Su, PhD, associate professor and director of data management services in the department. Those points could include how tumors communicate with surrounding cells, where patients live near health care facilities, the order in which medical events occur, and comparisons between patients with a common disease. Graphs can also represent knowledge, guide AI models and help them gain new knowledge.

More complex studies can combine all these variables to find unexpected links and evaluate treatment options. Right now, Jing said, technology is revolutionizing care. "Previously, many models only gave a patient some predictions about what would happen next," he said. "Now, we start using AI to find the best way to safely and dynamically manage the care of this patient to the best results."

AS BOTH ACADEMIA and industry race ahead with artificial intelligence, the School of Medicine is in a strong position. IU has powerful computers, such as the Big Red 200 supercomputer, and offers faculty access at a much-lower cost than many other major research centers.

"When I talk with my colleagues about the level of compute power that we have access to — and what they cost — they are just amazed," said Shaun Grannis, MD, MS, the Regenstrief Professor in Medical Informatics and a professor of family medicine at IU.

Spyridon Bakas, PhD, left the University of Pennsylvania about a year ago to become IU School of Medicine’s director of computational pathology. He’s been harnessing the power of big data in research related to multiple forms of cancer. At IU, he’s been impressed by the high-performance computing power, the expertise of the people engaged in the technical work, and the collaborative culture between people from different disciplines, and between clinicians and researchers. “We have a lot of competence at IU,” he said, “but we really don’t advertise it.”

AI models are also hungry for data. To improve their accuracy, they require as much data as possible, representing diverse patient populations. For that reason, IU researchers collaborate with other universities, medical schools, and research consortiums to ensure access to large, combined data resources. Then, there is the school's partnership with health systems around Indiana.

"IU offers unfettered access to the data of a third of all patients in the state, and we have a strong presence in data science and data analytics,” said Michael Feldman, MD, PhD, chair of the Department of Pathology and Laboratory Medicine and the Manwaring Professor of Pathology and Laboratory Medicine at the IU School of Medicine. "There's a huge opportunity to move this science forward."

The promise of the technology, already being realized, is to help researchers work faster and expand the scope of their gaze to thousands of patients across the country or around the world. It offers to do tedious jobs much quicker — sometimes converting weeks of human work into days, and days into hours, or minutes. It holds the promise to clinicians of less documentation and better follow-up care. For radiologists, a second read on images, reducing the likelihood the vital stuff isn't overlooked.

This raises the question: Will artificial intelligence make humans in medicine obsolete? The consensus at IU School of Medicine is a resounding no.

Huang, the chair in bioinformatics, said it's not a question of whether AI is better than humans. "If you ask AI to read a thousand papers, of course, it's going to be better — it's a computer," he said. "But on the other hand, if you really want AI to help you understand that and provide you something really explainable and accurate and understandable, then no."

Dana Mitchell, MD, MS, an assistant research professor of pediatrics, is actively training cutting-edge AI software to identify tumor tissue within the peripheral nervous system. When she's done, the program allows her to analyze a much larger number of samples more quickly than would have been possible if she or one of her technicians had to review such a large quantity of slides themselves.

ANOTHER WAY TO look at it is like a plane on autopilot, said Kevin L. Smith, MD, an assistant professor of clinical radiology and imaging sciences who leads the clinical AI program in radiology at IU Health. “Even as the autopilot is engaged, the human pilot validates that the system works as it should.”

And there is still plenty of work to do.

Algorithms are not built in isolation. They are trained on and analyze data extracted from a real and tactile world. It also means that tools intended to be objective are not entirely free of bias. As Grannis notes, AI models should behave in ways we expect and with reliability in results that can be replicated. Since much of the work is built on patient records, close attention must also be paid to confidentiality and privacy.

There are even issues with health disparities, Huang said. As AI is deployed in clinical settings, how widely will it be used, and which patients will have access to its benefits? Academic health centers and hospitals in wealthy communities are already employing it. But what about health care settings in lower-resource rural areas or low-income countries, such as Kenya? AI has the potential to bridge these gaps — giving radiologists in remote areas an AI backup read — or, in AI’s absence, to make disparities worse.

Pollok, who met Tyler Trent before his death and is now one of the caretakers of his tumor data, agrees. A biologist by training and experience, she claims to be learning bioinformatics — and AI — as she goes, despite evidence she's well-skilled. But Pollok is hopeful that, before she retires, her lab can push two potential therapies to clinical trials, including one for Tyler's disease, osteosarcoma.

If it happens, it's likely that it would have been made possible with the help of artificial intelligence.

Laura Gates and Matthew Harris contributed to this story.