Integrating AI in clinical decision-making

Large language models have achieved impressive results in medical reasoning and diagnostic tasks, but their role in healthcare remains largely confined to standalone applications. Whether they can be integrated into clinical practice, including managing patient care directly within electronic health records (EHRs), remains unproven. In a publication in Nature, Ferber et al. developed MIRA (Medical Intelligence for Reasoning and Action), an autonomous agent designed to operate in a sandboxed EHR environment, which showed performance that was comparable to or above that of physicians on diagnosis and treatment quality while adhering to clinical guidelines.

The authors evaluated MIRA using simulated patient cases derived from the MIMIC-IV database. The system was able to collect patient histories, order and interpret diagnostic tests, generate differential diagnoses, and propose treatment plans. MIRA’s performance was notable, achieving diagnostic accuracy comparable to or exceeding that of the physician cohorts. The AI agent demonstrated high alignment with established clinical practices, particularly in surgical procedures and medication management. MIRA’s therapeutic decisions showed greater concordance with reference clinical datasets and medical guidelines than those made by physicians. The study also highlighted MIRA’s ability to provide safe patient-level prescriptions and appropriate emergency department admission or discharge recommendations. Importantly, MIRA maintained stable performance across various perturbation scenarios, indicating robustness to different patient traits.

This is a preview of subscription content, access via your institution