Researchers at Stanford University have developed benchmarks to evaluate the effectiveness of AI agents in health care that order tests, suggest medications, retrieve patient information and perform other administrative tasks. The study, published in NEJM AI, found that Claude 3.5 Sonnet v2 had the highest success rate at 69.67%, while Mistral v0.3 had the lowest at 4%, and newer models performed better than older ones.

AI tools developed for mental health support may not recognize suicidality while stigmatizing some conditions in ways that create barriers to care, researchers have found. Psychiatrist Ravi Hariprasad suggests steps mental health clinicians can take, starting with asking patients about their AI use, educating them about safe and unsafe practices, and helping them set boundaries.

The Social Security Administration currently takes more than 200 days to process initial disability claims, partly due to delays in obtaining medical records, but the Trusted Exchange Framework and Common Agreement can automate and expedite the process. TEFCA's underlying technical specifications and the Qualified Health Information Network Technical Framework will be updated to support government benefits determinations.

An in-house deep learning AI algorithm developed using data from the National Lung Screening Trial predicted lung nodule malignancy as well as the Pan-Canadian Early Detection of Lung Cancer model, with a 40% lower false-positive rate, researchers reported in Radiology. The AI model performed better than the PanCan model with indeterminate nodules and with malignant nodules that were size-matched with benign nodules. "Deep-learning algorithms can assist radiologists in deciding whether follow-up imaging is needed, but prospective validation is required to determine the clinical applicability of these tools and to guide their implementation in practice," said lead author Noa Antonissen.

A study published in JAMA Network Open suggested that a clinical decision support tool could improve evidence-based testing for pediatric diarrhea and enhance communication between clinicians and parents about the causes of illness. However, parents were wary about the use of such tools, fearing over-reliance on algorithms and a lack of patient-centered care.

A data collection method called AEquity uses a learning curve approximation to identify and decrease bias in health data sets. The tool reduced bias by up to 96.5% in chest radiograph data and up to 80% in mortality prediction data, according to a study in the Journal of Medical Internet Research. Researchers say the tool could improve fairness in health AI if used during algorithm development and pre-deployment audits.

As companies increasingly deploy AI agents powered by large language models, new risks such as deception and data loss are emerging. To mitigate these risks, experts recommend setting strict limits and guardrails on AI behavior, continuously monitoring AI actions and preparing robust incident response plans.

An AI voice-based agent helped older adults with hypertension accurately report home blood pressure measurements, according to a study presented at the American Heart Association's Hypertension Scientific Sessions. The AI agent was associated with a 17% improvement in Medicare Advantage Stars ratings for blood pressure control. The study also found high patient engagement and satisfaction.

A report published in the Journal of the American Medical Association describes a national data collection and reporting tool that tracks measles cases by county across the US. Researchers say county-level data is needed because many measles outbreaks are localized and the CDC tracks cases at the national and state levels. The dashboard documents a recent surge in measles cases that coincides with a decline in vaccination coverage since 2020.

Health care professionals using EHRs are invited to participate in the Fall 2025 25x5 TrendBurden Pulse Survey. The brief, 2-minute survey tracks national documentation burden trends and provides critical evidence to policymakers and health system leaders. This semiannual survey ensures health care worker voices are represented in discussions about EHR improvements. Participants who completed the spring survey are encouraged to take it again to help researchers track changes over time. The survey is open to all health care professionals. Take the survey today.

AHIC addresses the growing need for standardized competency recognition among health informaticians working across health care delivery, public health and consumer health. Professionals who pass the 150-question exam earn the prestigious AMIA Certified Health Informatics Professional (ACHIP) designation. AHIC is designed for mid-level to advanced health informatics professionals from diverse backgrounds, including medicine, nursing, pharmacy, public health, computer science and more. Begin your health informatics certification journey.