The landscape in LLM and healthcare is expanding faster than climate change. LLMs are advancing day by day. There are many wannabees, and most of the offerings are not verified, nor certified by the FDA. None of them are HIPAA compliant.
Health in ChatGPT
We’re starting to roll out Health, a dedicated space in ChatGPT for health and wellness conversations. You can securely connect medical records, Apple Health and supported wellness apps so answers can be grounded in your own data. Health is designed to help you navigate medical care, not replace it. Health conversations, memory, and files are kept separate from the rest of ChatGPT and are not used to train our foundation models.
Health appears as a new space in the ChatGPT sidebar. If you ask a health-related question in a regular chat, ChatGPT may also suggest continuing in Health. If you’re interested in getting access as it becomes available, you can sign up for the waitlist. We’re starting by providing access to a small group of early users to learn and continue refining the experience. At launch, Health is available on web and iOS, with Android coming soon to ChatGPT Free, Go, Plus, and Pro users in supported countries (excluding the EEA, Switzerland, and the UK).
It is not HIPAA compliant. It has not been verified as to accuracy.
Accuracy & Reliability Issues
Hallucinations: Can generate false information or fake sources that sound credible.
Inconsistent Accuracy: Clinical accuracy varies widely (20-95%) and is unreliable for complex tasks or diagnoses, notes Healthgrades Health Library.
Outdated Knowledge: Its training data often stops in 2021, missing recent medical advancements.
Not to be outdone Anthropic CLAUDE jumps on the bandwagon with Anthropic announcing Claude for Healthcare following OpenAI’s ChatGPT Health revealOn the heels of OpenAI’s ChatGPT Health reveal, Anthropic announced on Sunday that it’s introducing Claude for Healthcare, a set of tools for providers, payers, and patients.
Like ChatGPT Health, Claude for Healthcare will allow users to sync health data from their phones, smartwatches, and other platforms (both OpenAI and Anthropic have said that their models won’t use this data for training). But Anthropic’s product promises more sophistication than ChatGPT Health, which seems as though it will be more focused on a patient-side chat experience as it rolls out gradually.
Though some industry professionals are concerned about the role of hallucination-prone LLMs in offering clients medical advice, Anthropic’s “agent skills” seem promising.
Claude has added what it calls “connectors” to give the AI access to platforms and databases that can speed up research processes and report generation for payers and providers, including the Centers for Medicare and Medicaid Services (CMS) Coverage Database; the International Classification of Diseases, 10th Revision (ICD-10); the National Provider Identifier Standard; and PubMed.
Anthropic explained in a blog post that Claude for Health could use its connectors to speed up prior authorization review, the process in which a doctor must submit additional information to an insurance provider to see if it will cover a medication or treatment.
“Clinicians often report spending more time on documentation and paperwork than actually seeing patients,” Anthropic CPO Mike Krieger said in a presentation about the product.
For doctors, submitting prior authorization documents is more of an administrative task than something that requires their specialized training and expertise. It’s something that makes more sense to automate than the actual process of administering medical advice … though Claude will do that as well.
Specialized Medical LLMs
Med-PaLM 2 / MedLM: Developed by Google, Med-PaLM 2 was the first AI system to reach human-expert levels on USMLE-style questions. It is part of the MedLM family of foundation models fine-tuned specifically for healthcare applications such as summarizing patient records and answering medical queries.
BioGPT: A Microsoft-developed model trained on large-scale biomedical literature. It excels in natural language processing tasks like relation extraction (identifying links between genes and diseases) and text generation for biomedical research.
BioMistral: An open-source collection of pre-trained models for the medical domain, built upon the Mistral architecture to provide accessible, high-performance medical AI.
AMIE (Articulate Medical Intelligence Explorer): A Google research model specifically optimized for clinical conversations and diagnostic reasoning.
GatorTron: An LLM developed specifically to process electronic health records (EHRs), focusing on semantic similarity and clinical question answering.
BioMedLM: A model focused on biomedical domain-specific question answering.
Task-Specific Models
Radiology-Llama2: Specialized for generating and analyzing radiology reports.
BioBART: Optimized for medical dialogues, summarization, and named entity recognition (NER) in healthcare data.
DeID-GPT: Designed specifically for the de-identification of patient data to ensure privacy and compliance.
ChatCAD: A model focused on computer-aided diagnosis to assist clinicians in interpreting medical images and data.
Healthcare-Specific Platforms
Amazon Bedrock: While a general platform, it provides access to specialized foundation models used by healthcare organizations like Fujita Health University to automate discharge summaries.
Microsoft Healthcare Agent Service: A suite of tools built on LLMs to power medical virtual assistants and streamline clinical documentation.
Specialized Medical Imaging LLMs
In 2026, the landscape of medical imaging is being transformed by Multimodal Large Language Models (MLLMs) and Vision-Language Models (VLMs). Unlike standard text-based LLMs, these models integrate image encoders (such as Vision Transformers) with language decoders to interpret 2D and 3D medical data alongside clinical text.
Leading Multimodal Medical Models
RAD-DINO: Developed by Microsoft Research and the Mayo Clinic, this foundational model is trained solely on biomedical imaging. It analyzes chest X-rays and associated clinical reports to automate report generation and verify medical device placements.
Pillar (Pillar-0): An open-source model released by UC Berkeley and UCSF in late 2025. It is specifically designed to interpret 3D volumes directly (CT/MRI) rather than just 2D slices, allowing it to recognize hundreds of conditions from a single scan.
MedGemma Multimodal (27B): A high-capacity open model from Google designed for interpreting complex multimodal data and longitudinal electronic health records (EHRs).
MedSigLIP: A specialized image and text encoder also from Google, recommended for structured imaging tasks like classification and retrieval.
Medical VLM-24B: A major release from John Snow Labs tailored to medical specialties, offering high standards of accuracy for clinical visual tasks.
LLM Radiology: AI-Powered Reporting and Diagnosis
Localization Lens Boosts Medical Vision-Language Models at ...
Specialized Emerging Models
RetinaVLM: A domain-specific model developed to handle specialist clinical knowledge in ophthalmology, outperforming general models in creating detailed imaging reports for managing conditions like AMD.
ClinicalBLIP: A model that utilizes a multistage fine-tuning strategy to improve the generation of radiology reports from medical images.
TransUNet / RadFormer: Earlier transformer-based models that have evolved to perform semantic segmentation while simultaneously generating natural language descriptions of the segmented regions.
Key Capabilities in 2026
Automated Report Drafting: Drafting preliminary radiology reports by comparing current scans with prior images and patient history.
Visual Question Answering (VQA): Allowing clinicians to ask specific questions about an image, such as "What is the size of this lesion compared to the previous scan?".
Intelligent Orchestration: Moving from simple interpretation to "orchestrating intelligence," where AI links imaging findings with pathology, genomics, and clinical notes.
Multimodal Large Language Models in Medical Imaging
Aug 8, 2025 — MLLMs, also known as large multimodal models (LMMs), represent a significant evolution in medical AI [6]. Their core ca...
:: KJR :: Korean Journal of Radiology
MedGemma: Our most capable open models for health AI ...
Jul 9, 2025 — In May of this year, we expanded the HAI-DEF collection with MedGemma, a collection of generative models based on Gemma...
Google Research
Large Language Models in radiology: A technical and clinical ...
It is worth noting that LLMs in radiology might also benefit from multimodal training: pairing images with text. While traditional...
ScienceDirect.com
Disclaimer:
People are already relying on LLMs for medical advice. OpenAI said that 230 million people talk about their health with ChatGPT each week, and there’s no doubt that Anthropic is observing that use case as well.
Of course, both Anthropic and OpenAI warn consumers that they should see healthcare professionals for more reliable, tailored guidance.
What are some emerging LLMs that can analyze medical images?
No comments:
Post a Comment