In partnership with
Introducing the first AI-native CRM
Connect your email, and you’ll instantly get a CRM with enriched customer insights and a platform that grows with your business.
With AI at the core, Attio lets you:
Prospect and route leads with research agents
Get real-time insights during customer calls
Build powerful automations for your complex workflows
Join industry leaders like Granola, Taskrabbit, Flatfile and more.
Hi!
Last week, we talked about OpenAI’s Health and Healthcare release.
Around the same time, Anthropic also introduced Claude for Health.
Their product is designed to improve workflows from clinical trial management to regulatory operations, all while being HIPAA-compliant.
Anthropic also supports patients. Users can connect their health data, get information explained clearly, and summarize their medical history and lab results. (similar to ChatGPT Health). They’re tackling the administrative nightmare.
Have you tried either? Let me know your thoughts.
On another hand, I’m adding a new section with great papers worth reading if you have extra time: 🧪 Research Signals. Let me know what you think.
Now, let’s dive into today’s issue.
🤖 AIBytes
This international retrospective study tested if a medical AI model can improve over time by learning sequentially from new hospitals, and whether this improves real-world performance.
🔬 Methods
Clinical task: Measuring endotracheal tube (ETT) position relative to the carina on chest X-rays.
Data:
2,313 ICU chest radiographs
23 hospitals across 12 countries and 5 continents
100 images per hospital (50 for training, 50 for testing)
Model: Convolutional neural network (CarinaNet)
Training strategies compared:
No adaptation (original model)
Single hospital fine-tuning
Continual learning (sequential fine-tuning across hospitals)
📊 Results
Continual learning:
Outperformed the original model at all hospitals.
Outperformed single-hospital fine-tuning at 21 of 22 hospitals.
Model performance improved progressively as data from more hospitals were added.
🔑 Key Takeaways
Continual learning improves accuracy using small local datasets.
Performance gains were achieved without sharing patient data.
This approach may be practical for hospitals with limited infrastructure.
🔗 Chen E, Saenz A, et al. International retrospective observational study of continual learning for AI on endotracheal tube placement from chest radiographs. NEJM AI. 2025;3(1). doi:10.1056/AIoa2500522
Researchers reviewed how AI is transforming digital twins. From static simulations into adaptive systems that can predict and intervene, they will eventually be able to act autonomously across medicine and other fields.
🔬 Methods
Framework proposed: Four sequential stages
Modeling: Digital twins are used to model patient-specific physiology using multimodal data.
Mirroring: The systems mirror the patient’s state over time by continuously updating simulations.
Intervention: Some models are designed to simulate or suggest interventions.
Autonomous management: Independently select, adapt, and execute clinical management actions over time.
Domains reviewed: 11 application areas, including healthcare, biology, robotics, and aerospace.
🔑 Key Findings
AI has shifted digital twins from passive models to learning systems capable of:
Predicting system behavior
Detecting anomalies earlier
Simulating “what-if” clinical scenarios
Physics-informed AI improves:
Generalization across conditions
Interpretability of model behavior
Speed
Generative AI and LLMs enable:
Agent-based reasoning
Closed-loop decision support, where systems update based on feedback.
Major challenges remain unresolved:
Scalability
Explainability
Trustworthiness
Ethical and safety governance
🩻 Why This Matters
Digital twins could support personalized care, early risk detection, and treatment simulation.
Digital twins are evolving toward adaptive clinical systems, not just simulations.
Most healthcare applications remain research-stage.
Autonomous decision-making is not ready for clinical deployment.
🔗 Zhou R, Chen D, Jia Z, et al. Digital Twin AI: Opportunities and Challenges from Large Language Models to World Models. arXiv. 2026. doi:10.48550/arXiv.2601.01321
🦾TechTools
Combines LLMs with structured medical knowledge to support clinical workflows.
Assists with documentation, reasoning, and information retrieval.
HIPAA-compliant platform that also offers AI-assisted health assessments and connects patients with licensed clinicians.
Reads the full medical record (notes, outside records, scanned documents) and turns it into clinical insight inside the EHR.
Supports real workflows: emergency medicine, inpatient care, perioperative teams.
Focuses on reducing cognitive burden while improving quality metrics and documentation.
A no-code platform for creating interactive learning and training content.
You can build simulations, tutorials, and educational material.
Useful for medical education.
🧬AIMedily Snaps
AI is speeding into healthcare. Who should regulate it? Harvard Gazette (Link).
Stanford report: AI-driven insurance decisions raise concerns about human oversight (Link).
Now is the time: turning the promise of AI into health for all (Link).
Google released a new generation of medical image interpretation with MedGemma 1.5 and medical speech to text with MedASR (Link).
OpenAI acquired Torch, that unifies lab results, medications, and visit recordings (Link).
Google removes some of its AI summaries after users health was put at risk (Link).
The Mount Sinai Cancer Center has launched an AI platform to connect cancer patients to clinical trials (Link).
🧪 Research Signals
Discovery of predictive biomarkers for cancer therapy through computational approaches (Link).
ChatGPT-guided translation of pediatric emergency department discharge instructions (Link).
Effects of Artificial Intelligence Recognition–Based Telerehabilitation on Exercise Capacity in Patients With Hypertension: Randomized Controlled Trial (Link).
Machine learning in the development and application of patient-reported outcome measures (PROMs) for surgical patients: a systematic review (Link).
Digital biomarkers for brain health: passive and continuous assessment from wearable sensors (Link).
🧩TriviaRX
What was one of the first real clinical uses of machine learning in medicine?
A) Cancer genomics
B) ECG interpretation
C) ICU mortality prediction
D) Radiology image classification
Now, it’s time to check if you got the right answer from last’s week TriviaRx:
✅ B) Radiology
In one large cross-sectional study of over 900 FDA-approved AI/ML devices, radiology accounted for 76.6% of devices.
That’s it for today.
As always, thank you for being here.
If this issue was useful, feel free to share it with a colleague. Thank you!
Until next Wednesday.
Itzel Fer, MD PM&R
Forwarded this email? Sign up here
P.S. Enjoying AIMedily? 👉 Write a review here (it takes less than a minute).
