Post by ARISE

🔍 Can clinical AI "see" as well as it "reads"? A new study in Nature Communications, authored by ARISE's own Thomas Buckley, Peter Brodeur, MD, MA , Adam Rodman, Arjun (Raj) Manrai, and colleagues evaluated 8 vision-language models including o3 and GPT-5 on 1,090 multimodal medical cases. Here's what they found: 1. Accuracy tracks the text, not the image. Model performance increased steadily with the amount of informative text in a case. When the text was already robust, adding the image did little to improve accuracy. In long clinicopathological cases, where more text is available, the image sometimes actually lowered performance! Human physicians, by contrast, didn't improve with more text. 2. That strength is also the weakness. When researchers inserted a misleading text-based clinical vignette, o3's accuracy on cases it had previously nailed from the image alone fell from 84% to 28%. The models rarely pinpointed the mismatch, it just reinterpreted the image to fit the wrong text. Well worth a read if you work at the intersection of AI and medicine. 🔗 [link in the comments below] #ClinicalAI #HealthcareAI #MultimodalAI #MedicalImaging #AIinMedicine #MachineLearning #DigitalHealth #FoundationModels