Isaac "Zak" Kohane, MD, PhD, the inaugural chair of Harvard Medical School's Department of Biomedical Informatics, discussed artificial intelligence during a plenary session at IDWeek 2023.
At the opening plenary of IDWeek 2023, Isaac "Zak" Kohane, MD, PhD, the inaugural chair of Harvard Medical School's Department of Biomedical Informatics, engaged infectious diseases experts by discussing the potential of medical AI while also highlighting its challenges.1
Kohane, a physician-scientist renowned for his work at the intersection of artificial intelligence and precision medicine, first showcased the rapid capabilities of AI with a case study from the Undiagnosed Diseases Network (UDN).
The UDN program, aimed at providing answers for patients and families affected by undiagnosed diseases, performed whole exome sequencing on a 4-year-old boy who had developed unexplained dystonia in 2021. In this "old-school AI" approach, they employed machine learning to identify guanosine triphosphate (GTP) cyclohydrolase I deficiency from a list of potential variants. Upon initiating folinic acid, L-DOPA, and 5-hydroxytryptophan, the boy regained motor function.
One year later, by simply providing OpenAI's Generative Pre-Trained Transformer-4 (GPT-4) with a UDN case summary and a list of whole exome sequencing variants, the model suggested POLR3A as an "interesting" example. This finding was swiftly verified through cell models.
The rapid evolution of AI in medicine is further illustrated by the progress of these models in answering US Medical License Exam (USMLE)-style questions. Previously, AI models struggled to pass USMLE-style questions, but by late 2022, Google's MED-PaLM achieved passing scores. A few months later, its successor, MED-PaLM 2, scored even higher.
Despite the complexity of medicine not always aligning with the constraints of earlier AI breakthroughs, Kohane discussed the progress in medical AI, emphasizing that "It is getting better because we're getting more data." He later demonstrated how AI can swiftly generate detailed echocardiography reports and expedite the identification of clinical criteria for severe SARS-CoV-2 infection during the early days of the pandemic. He even employed GPT-4 to analyze CDC data on antimicrobial resistance, conduct taxonomy analyses, and visualize resistance rates geographically, all without writing a line of code.
While these AI advancements are impressive, Kohane also highlighted the challenges associated with current AI tools, including the potential for outdated information, errors, and even fabricated data. In a well-known misapplication, a COVID-19 ICU prediction model trained on one patient population failed to predict the severity of later SARS-CoV-2 variants.
Despite these limitations, Kohane emphasized that, as patients gain access to these tools, they will rapidly adopt them. Looking ahead, he envisions a global "sea change in healthcare with AI" that will outpace both regulation and education. As AI accelerates clinical and basic research and transforms healthcare delivery, Kohane recommends that institutions allocate resources for data and computation and develop expertise in quantitative data analysis to ensure that healthcare providers maintain a prominent role in this evolving landscape.