During the IDWeek 2023 Opening Plenary on Wednesday afternoon, Isaac Kohane, MD, PhD, chair of the department, Harvard Medical School, Boston, Massachusetts, described artificial intelligence (AI) fundamentals applied to medical practice, identified current and potential AI applications in ID, including potential pitfalls, and analyzed its role in improving diagnosis and treatment planning.

Dr. Kohane related the story from 2021 of a boy with worsening dystonia, characterized by not walking or talking. Utilizing the Undiagnosed Diseases Network and machine learning (Dr. Kohane is the Principal Investigator of the network’s coordinating center), physicians were able to identify that he had GTP cyclohydrolase I deficiency and administered a “cocktail” of L-dopa, folinic acid, and 5-hydroxytryptophan. “Low and behold,” he said, “in a few weeks this patient was walking and talking.”

“This is ‘old school’ AI as of 2021,” he said. To illustrate how fast AI capabilities are growing, he contrasted this success with a more recent example in which an open-source AI database, Generative Pre-trained Transformer 4 (GPT-4), quickly helped guide treatment of a patient suffering from leukodystrophy, a genetic disorder affecting white matter in the brain.

These successes notwithstanding, the more likely impact on medicine, Dr. Kohane believes, will be on the business side. Billing, reimbursement, and prior authorizations, he expects, are where AI might help cut into healthcare costs. Citing the current primary care physician shortage, he also discussed AI’s potential role in filling these voids by helping train physician assistants and nurse practitioners. “There are fewer applicants than there are [primary care physician] slots,” he observed.

Despite the promise of AI in healthcare, Dr Kohane said, “The future of AI in healthcare is very muddy,” he said. He explained that the integrity of the data that AI utilizes is critical to ensure its acceptance and wide adoption. He recommended that healthcare practitioners adopt a “trust but verify” approach. “But that’s easier said than done,” he said. “Once you put a default dose in the order entry system, they [clinicians] use that default dose 95% of the time. How do we actually make sure that clinicians stay awake and not just let AI autonomously do things that it may not do well because it does so many things?”