In partnership with
A newsletter on the latest in AI for healthcare.
Welcome back,
The learning healthcare AI company news today is Tempus ECG-PH receiving FDA 510(k) clearance. Its health AI product analyses routine 12-lead ECGs to flag signs of pulmonary hypertension.
In a research study published in Nature Medicine, an LLM clinical decision support system was evaluated in ED, the result was that clinicians used the AI system less with time, despite the LLM producing safe and clinically appropriate outputs.
Also today, ClinicalThought-AI-8B is our chosen model to tinker with / read about, it offers an open, locally deployable medical reasoning model! (exciting).
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.
SUMMARY
Top Research Paper
An LLM evaluated in the real-world emergency department found clinically appropriate outputs but rapidly declining clinician use.
Top AI News
Tempus ECG-PH received FDA 510(k) clearance to flag signs associated with pulmonary hypertension from routine 12-lead ECGs.
Top Model
ClinicalThought-AI-8B is an open 8-billion-parameter medical reasoning model that can run locally for tasks such as differential-diagnosis support.
Bedside Bets
Startup rounds, deals, and moves.
Arintra (a YC company) raises $25M to scale autonomous medical coding inside existing EHR workflows: Arintra uses LLMs and clinical knowledge graphs to turn unstructured charts into billing codes, with integrations including Epic and Athena health.
Pulse Check
Quick reads across health AI.
FDA seeks input on regulating generative AI medical devices: The FDA is considering rules for clinical risk, autonomy, premarket testing and postmarket monitoring. Comments are due 19 October 2026.
Mercy expands Aidoc imaging AI across 50+ hospitals: Mercy now uses 13 AI tools to analyse about 2.5 million cases annually, prioritising findings such as pulmonary emboli, brain bleeds and fractures.
AWS launches student rewards on AWS Builder Center: Students earn points and rewards through learning activities, challenges and community contributions. Sign up to build practical AWS skills, track your progress and access student-focused opportunities.
TOP PAPER
👨🏾⚕️ Prospective evaluation of a large language model clinical decision support system in the ED
Source: Nature Medicine · 19 August 2026
An LLM-based clinical decision-support system was tested prospectively in a live tertiary emergency department. Across 1,138 patients over four weeks, clinicians rated 99 of 100 sampled outputs as clinically appropriate, but use fell from 68% to 30% and length of stay was unchanged.
The result highlights a practical challenge for clinical AI: even useful tools may be ignored if they add another interface or task during periods of high cognitive load. The strongest systems may be those that fit invisibly into existing workflows and reduce clinicians’ cognitive burden.
Research Question
Can an LLM-based clinical decision-support system be integrated safely into emergency care while sustaining clinician use and improving operational outcomes?
Approach
1,138 patients were analysed over four weeks across two emergency department units.
The LLM connected to the electronic health record (EHR), generated structured histories and clinical insights, and was accessed through a retrieval-augmented chat interface.
Researchers measured adoption, safety, clinical appropriateness, length of stay and consultation time.
Results
Clinician adoption fell from 68% to 30%.
Use declined with each additional hour of a shift: OR 0.72.
Physicians were more likely to use the system for radiology consultations: OR 2.98.
No adverse events were detected.
Expert reviewers rated 99 of 100 sampled outputs as clinically appropriate.
Length of stay was identical at 4.9 hours in both groups.
Consultation time fell by 9.4 minutes, but not significantly (P = 0.077).
Caveats
Single-site, early-stage evaluation.
Safety assessment included only 100 sampled outputs.
Consultation findings may reflect selective adoption.
The study does not establish that the system is ready for clinical deployment.
Potential impact:
The main lesson may be about product design, accuracy and safety matter, but so does cognitive fit. Clinicians working under pressure may not use a tool that requires another interface or behaviour change.
TOP NEWS
Tempus can now use a routine ECG to flag patients who may have pulmonary hypertension
Source: Tempus / Business Wire · 24 August 2026
Tempus has received FDA 510(k) clearance for Tempus ECG-PH, which analyses routine 12-lead ECGs for signs associated with pulmonary hypertension.
The tool is intended for symptomatic patients aged 40 and older with no known pulmonary hypertension. It returns a binary result based on ECG patterns linked to mean pulmonary artery pressure above 20 mmHg. It is not a standalone diagnostic or for serial monitoring, and results should be interpreted alongside other clinical information.
Why it matters:
Tempus is using data already collected in routine care to add a clinical signal without requiring a new device or workflow. ECG-PH is the company’s third FDA-cleared cardiovascular AI product.
NEWS SOURCE >»
TOP AI MODEL
ClinicalThought-AI-8B brings local medical reasoning to an 8B model, for example generating a structured differential from a clinical case
Source: Hugging Face · August 2026
ClinicalThought-AI-8B is an open 8-billion-parameter medical reasoning language model designed for clinical reasoning, differential diagnosis, case consultation and medical education.
It is based on IBM Granite 3.3 8B and is distributed in formats that make local deployment relatively straightforward.
What stands out:
8B parameters, small enough to be more practical for local development than many frontier medical models.
Apache-2.0 licence.
Based on IBM Granite 3.3 8B with an adapter-based medical fine-tune.
Includes SafeTensors and GGUF weights plus LoRA adapters, training data, inference scripts and training logs.
Supports deployment routes including llama.cpp, vLLM, Ollama and Docker.
BEDSIDE BETS
Bedside Bets
Startup rounds, deals, and moves in healthcare AI.
YC company Arintra raises $25M Series B to automate medical coding: Its AI extracts billing codes from unstructured EHR data, with reported results including 5%+ revenue uplift, 12% fewer accounts-receivable days and 43% fewer coding-related denials.
If this newsletter was forwarded to you, subscribe here or see more
NEWSLETTER BY:
Dr Ezekiel Dinama
MD and PhD Researcher at Cambridge University applying physics-informed ML/AI to neurophysiological research.