JMIR articles address AI clinical decision-making and health care worker burnout

JMIR Publications released two feature stories in its News and Perspectives section. Shalini Kathuria Narang's "Can Humanlike Reasoning Be Replicated in Large Language Models for Clinical Decision-Making?" and Sara Novak's "How Health Care Workers Can Manage Digital Fatigue" offer complementary looks into the capabilities of artificial intelligence in diagnostics, and the real-world exhaustion faced by medical professionals managing digital systems.

Can LLMs match physicians in clinical reasoning?

In "Can Humanlike Reasoning Be Replicated in Large Language Models for Clinical Decision-Making?", Narang covers a recent study comparing the diagnostic reasoning of OpenAI's o1 model against physicians. Narang reports that the model matched or exceeded human performance across three stages of care-triage on arrival, first contact with a physician, and upon admission-with the widest performance gap occurring at initial ER triage where available information was most limited. Adam Rodman, hospitalist and one of the researchers on the study, notes that the results validate the diagnostic performance of the models, but do not mean the system is ready to be deployed independently.

While the LLM excelled at integrating text-based information, actual clinical practice relies heavily on nontext inputs like visual and auditory cues gathered during physical examinations. Rodman points out that while LLMs are excellent at synthesizing curated data or collecting verbal information, they cannot replace a physician's ability to physically examine a patient, hear the hesitation in their voice, or integrate messy information from multiple uncurated sources.

Rather than replacing doctors, writes Narang, the future of AI in medicine requires collaborative integration and careful evaluation. The researchers highlight the need for prospective trials in real-world settings to evaluate newer multimodal models safely. According to Rodman, a promising application for this technology is acting as a second opinion to catch diagnostic errors before they occur, alerting doctors if they might be heading in the wrong direction.

Confronting the digital workload

Novak investigates a paradoxical phenomenon which has emerged as clinical processes increasingly integrate with digital tools: digital fatigue. "How Health Care Workers Can Manage Digital Fatigue" explores how, despite the benefits of automation and increased access in clinical care, many health care workers find themselves struggling to keep up with the continuous demand to manage digital interfaces and respond to redundant alerts. Hearing from expert sources-physician Hassan Bencheqroun, and digital fatigue researchers Rachel Hoopsick and Audrey Hai-Novak outlines how, as the prevailing fee-for-service health care system already limits the time providers can spend with each patient, the additional administrative burden of navigating complex platforms creates a compounding cycle that drives clinician burnout. 

Novak outlines strategies for institutions and individuals to manage these risks. Experts urge healthcare systems to streamline digital workflows by reducing low-value prompts, such as warnings for non-life-threatening allergies, and eliminating redundant alerts that often go ignored. Restructuring tasks into team-based systems-where responsibilities like inbox management and medication refills are actively shared-can prevent the accumulation of after-hours administrative work. 

Ultimately, it's up to health care institutions to recognize digital tasks as an official part of the daily workload, ensuring adequate time and training are provided during patient assignments-but on a personal level, health care workers can schedule digital detox breaks and delay email deliveries until working hours to protect their recovery time. "Digital fatigue needs to be taken seriously," writes Novak, "like you would any other occupational risk."