The Vibe
Multimodal AI is getting its first real stress test against board-certified radiologists, and the results are mixed at best [1]. While models claim human-level performance on cherry-picked datasets, head-to-head comparisons with subspecialty experts reveal significant gaps in diagnostic reasoning and accuracy [2]. The rush to deploy AI across medical specialties may be outpacing our ability to validate these tools where precision matters most.
Research
•Multimodal large language models achieved only 65% accuracy interpreting neuroradiology cases from RadioGraphics, significantly trailing neuroradiology subspecialists who scored 89% — the gap widened further on complex cases requiring spatial reasoning and pattern recognition [1]
•Three AI chatbots (ChatGPT, Grok, MANUS) showed inconsistent performance diagnosing oral pathology cases, with accuracy ranging from 42-78% depending on lesion type — none matched human pathologist performance on malignancy detection, raising serious questions about consumer AI health tools [2]
•Deep learning model distinguished primary from metastatic mucinous ovarian carcinoma with 94% accuracy from histopathology images alone, potentially eliminating need for expensive immunohistochemistry panels in ambiguous cases [3]
•Modified U-Net architecture improved biomedical image segmentation by 12% over standard approaches by replacing conventional downsampling with attention-based feature extraction — solves the long-range context problem that's plagued medical imaging AI for years [4]
Clinical Practice & Ops
•Labcorp expanded its PathAI partnership to deploy FDA-cleared digital pathology platforms across its U.S. anatomical pathology lab network — the first major commercial rollout of AI-assisted histopathology at scale [5]
•Nextech launched Cora Scribe, an AI documentation tool built specifically for specialty practices rather than primary care — addresses workflow differences that generic scribes miss in cardiology, orthopedics, and dermatology [6]
•FDA issued safety alert on Trividia blood glucose monitors after 114 serious injuries and one death over 12 years — highlights ongoing quality control issues in point-of-care diagnostic devices that patients rely on daily [7]
Policy & Regulatory
•FDA unveiled draft guidance for personalized CRISPR therapy approvals, creating a streamlined pathway for patient-specific genetic medicines — could accelerate custom treatments for rare diseases by years while maintaining safety standards [8]
•Opioid overdose deaths dropped 32% from 2023 to 2024 (79,358 to 54,045), driven largely by decreases in fentanyl-involved deaths — the first sustained decline since the crisis began [9]
Industry & Products
•Lilly added multidose Zepbound KwikPens to its direct-to-consumer platform after FDA approval, expanding self-pay access to the obesity medication — signals pharma's growing comfort with bypassing traditional distribution channels [10]
•OpenAI appointed Arvind KC as Chief People Officer to help scale the company and strengthen culture during rapid AI development — leadership moves that could shape how AI companies manage talent in the coming years [11]
Podcasts (Hot Takes)
•Julie Yoo at a16z argues healthcare is uniquely positioned to benefit from AI more than any other industry because historical tech underinvestment lets providers leapfrog directly to AI-native care models [12]. Bold claim, but healthcare's legacy system burden might actually prove she's right.
One to Watch
NEJM's secondary stroke prevention guidelines just dropped, emphasizing rapid diagnostic workup and mechanism-specific therapy selection [13]. With AI models now claiming stroke subtype classification from imaging, expect validation studies comparing algorithmic vs. clinical decision-making in stroke centers.