Prospective study involving 98 patients at a primary care clinic at Beth Israel Deaconess Medical Center in the US.

Before their urgent care visits, patients chatted with AMIE about their symptoms.

Primary care providers could review AMIE's summaries before seeing their patients.

Physicians monitored all conversations for safety.

AMIE's first diagnosis was correct in 56% of cases, increasing to 90% when considering its top seven diagnoses.

Among 44 providers who reviewed AMIE's output and completed surveys, 75% found it helpful for visit preparation.

In the same group, 57% said AMIE might have changed how they approached the visit.

No conversations required a safety stop. Supervisors identified one AI error and added clinical information in five cases.

AMIE showed promising results in gathering patient histories and helping physicians prepare for visits with real patients.

Patients and physicians generally had positive experiences with AMIE.

Every conversation was monitored by a physician.

The study was small and did not measure whether AMIE saved time or improved patient outcomes. Larger studies are needed.

Randomized controlled trial involving 2,378 adults across 24 hospitals in Tanzania.

AI group: Patients received immediate screening results and face-to-face counseling.

Standard care: Retinal images were reviewed by clinicians, with results communicated weeks later.

Researchers measured specialist clinic attendance within 8 weeks among patients who needed referral.

35.9% of patients in the AI group attended specialist care within 8 weeks, compared with 18.2% receiving standard care.

Among those who attended, patients in the AI group arrived sooner: 10.9 vs. 36.9 days, on average.

AI detected referable diabetic retinopathy with 91.6% sensitivity and 67.2% specificity.

The AI group had more false-positive referrals, potentially increasing specialist clinic workload.

AI-supported screening nearly doubled follow-up attendance and helped patients reach specialist care sooner.

The intervention combined AI with immediate results and counseling, so the benefit cannot be attributed to AI alone.

Despite the improvement, nearly two-thirds of patients in the AI group did not attend follow-up care.

Better screening is important, but access to specialist care and treatment remains essential to prevent vision loss.

OpenEvidence and Johns Hopkins Medicine partner to advance AI in medical education and clinical decision-making (Link).

MD Anderson researchers develop an AI tool that helps explain why immune cells enter some tumors but stay out of others (Link).

NIH joins an initiative to develop AI models that predict how human cells respond to disease and treatment (Link).

Google is using AI to accelerate disease research, from genetics to drug discovery (Link).

The Peterson Health Technology Institute (PHTI), a nonprofit that evaluates healthcare tech, releases a roadmap to accelerate the adoption of high-value clinical AI (Link).

UnityPoint Health expands its AI-powered medical record summaries across 18 hospitals to help physicians find important clinical information (Link).

npj: Researchers developed an AI model that uses routine chest X-rays to help identify patients who may have osteoporosis and need further screening (Paper).

Nature: In a study involving 13 patients with epilepsy, patient feedback helped reduce false seizure alarms by nearly 80% (Paper).

Nature: Researchers introduce MAGIC, an international initiative to evaluate medical AI across different healthcare settings and patient populations (Paper).

JACC: A trial involving 171 patients found that AI monitoring helped detect changes in heart failure, but did not significantly improve patient outcomes (Paper).

JAMIA: A study involving 220 physicians found that explaining how AI makes a diagnosis did not necessarily improve their trust in its recommendations (Paper).

npj: Patients and lung cancer survivors expect medical AI to be highly accurate, especially in avoiding missed diagnoses and false alarms (Paper).