1. Reasoning models improve primary diagnoses while generating more false positives
October 29 | Benchmarks & Evaluations
A Nature Communications study compared six matched pairs of reasoning and standard language models across 12 specialties using a long-context diagnostic benchmark. Reasoning models produced more accurate primary diagnoses, but generated over 34% more false-positive diagnoses and showed approximately 40% lower reproducibility. These benchmark findings do not establish clinical outcomes in prospective practice.
My Take: A correct leading diagnosis can coexist with an unsafe differential. Clinical evaluations should examine false positives and consistency as well as overall accuracy.
2. Multimodal medical AI shows promising imaging results across specialties
October 29 | Benchmarks & Evaluations
Researchers described EVLF-FM in Nature Medicine, a vision-language model tested across multiple medical imaging modalities. Development and testing involved more than 1.3 million samples, with external validation using 13,443 independent samples. Internal average diagnostic accuracy reached 89.5%, and physician ratings of explanations averaged 4.25 out of 5. The investigators acknowledged the need for prospective clinical evaluation.
My Take: Visual explanations may help physicians examine outputs, but explanation quality and diagnostic correctness need separate measurement.
3. Pew survey finds substantial discomfort with AI in diagnosis and treatment
October 29 | Governance & Risk
A Pew Research Center survey of 3,488 U.S. adults found that 60% would be uncomfortable with physicians using AI for diagnosis and treatment, compared with 30% who would be comfortable. Respondents generally favored human oversight for consequential medical decisions. The survey measures public attitudes rather than actual clinical AI performance.
My Take: Patient expectations should inform disclosure practices and decisions about the degree of physician review in clinical AI workflows.
4. Unauthorized AI use exposes patient information at Johns Hopkins
October 29 reporting | Governance & Risk / Safety
Healthcare IT News reported that an employee at Johns Hopkins Bayview Medical Center uploaded a spreadsheet containing patient information to an unapproved public AI service. The incident affected 3,182 patients and was discovered during an internal review in September. This is newly published coverage of an earlier privacy incident, and the report does not establish that the information was subsequently misused.
My Take: Policies need practical enforcement through approved alternatives, access controls, and staff training. The availability of public AI tools creates a persistent privacy exposure.