Is Google’s AMIE AI ready to replace the clinical intake interview? We break down the first real-world clinical feasibility study of conversational AI in primary care.

In this episode, we analyse a major prospective trial from Google Research and DeepMind testing the AMIE system on 100 urgent care patients. While the AI achieved zero safety stops and matched human doctors in diagnostic accuracy, a closer look at the workflow reveals significant hurdles. We explore the mechanics of clinical trust, why the messy reality of patient dialogue is the ultimate stress test, and why human doctors still beat AI on practical, cost-effective care plans.

Link to research report: https://arxiv.org/abs/2603.08448

DOI: https://doi.org/10.48550/arXiv.2603.08448

Link to associated blog post: https://research.google/blog/exploring-the-feasibility-of-conversational-diagnostic-ai-in-a-real-world-clinical-study/

Key Takeaways

• How conversational AI performs in a real-world primary care clinic without simulated patients.

• Why diagnostic accuracy doesn't automatically equal clinical trust, and why seeing the actual history-taking process is vital.

• The critical difference between an AI’s theoretical management plan and a human doctor’s practical, cost-effective clinical decision-making.

00:00 – Intro: A scenario of a patient completing an AI-led clinical interview.

00:32 – Study Introduction: Google’s AMIE (Articulate Medical Intelligence Explorer) powered by Gemini 2.5 Pro.

01:30 – Methodology: Real-world trials in a Boston primary care clinic with physician safety monitoring.

02:30 – Safety Results: Zero safety stops required during the trial encounters.

03:01 – Accuracy Results: Diagnostic performance compared to human primary care providers.

04:03 – Patient Feedback: Acceptance levels.

04:35 – Limitations: Issues with dialogue realism and the need for transcript transparency.

06:18 – Practicality Gaps: Why human doctors still outperformed AI on cost-effective management plans.

07:50 – Implementation Hurdles: Hardware limitations and demographic skews in the study.

09:31 – Governance & Validation: The importance of independent peer review (contrasted with Amazon).

10:51 – Future Outlook: Integration with Electronic Health Records (EHR) and multimodal (voice/image) capabilities.

13:34 – Conclusion: Summary of AMIE as a robust proof of concept for the future of patient journeys.

Clinical Governance & Educational Disclosure

This analysis is for educational and informational purposes only. It provides a technical review of AI in healthcare and does not constitute medical advice or treatment.

• Professional Accountability: If you are a healthcare professional, ensure your use of AI complies with local Trust policies and professional standards (GMC/NMC/HCPC).

• Evidence-Based Review: These views are my own and do not represent the official position of my University or Hospital Trust.

• Patient Safety: This video does not establish a doctor-patient relationship. Always seek the advice of a qualified healthcare provider regarding any medical condition.

Music generated by Mubert https://mubert.com/render

https://substack.com/@healthaibrief

#HealthTech #MedicalAI #GoogleHealth #PrimaryCare #ClinicalInformatics #DigitalHealth #DeepMind #FutureOfMedicine #EHR #MedicalInnovation

Are AI chatbots bypassing FDA regulation to deliver personalised medical advice? Explore the clinical and regulatory mechanics of the newly expanded Amazon Health AI.

This breakdown analyses the architecture of Amazon's agentic AI health assistant, now available across the primary Amazon app. By integrating nationwide Health Information Exchange (HIE) data, the system ingests electronic health records to provide tailored clinical guidance, explain lab results, and triage patients to One Medical providers. While the platform maintains strict HIPAA compliance for data security, the analysis investigates a critical regulatory gap: how software performing active clinical triage and personalized treatment routing currently operates outside traditional Software as a Medical Device (SaMD) definitions.

Link: https://health.amazon.com/health-ai/learn-more?ref_=hai_39_prk

Evidence of LLMs being unsafe at triage: https://youtu.be/BbB_FGu2uHk

Key Takeaways:

• Understand the multi-agent architecture of Amazon Health AI and how it integrates nationwide electronic health records directly into the consumer retail ecosystem.

• Differentiate between data security (HIPAA compliance) and clinical safety (FDA oversight), and why privacy alone does not guarantee algorithmic efficacy.

• Identify the regulatory blind spot allowing advanced LLMs to perform clinical triage and direct patient care pathways without traditional medical device classification.

00:00 – Intro: A scenario of a patient using the Amazon app for medical advice.

00:33 – Announcement: Amazon Health AI integration across the USA.

01:03 – System Architecture: How the agentic AI works.

02:18 – Safety & Ethics: Data security vs. clinical efficacy.

04:09 – Regulatory Issues: Lack of medical device status/FDA approval.

06:10 – Future Outlook: Benefits of modernizing healthcare access.

08:18 – Conclusion: Summary of potential and risks.

Clinical Governance & Educational Disclosure

This analysis is for educational and informational purposes only. It provides a technical review of AI in healthcare and does not constitute medical advice or treatment.

• Professional Accountability: If you are a healthcare professional, ensure your use of AI complies with local Trust policies and professional standards (GMC/NMC/HCPC).

• Evidence-Based Review: These views are my own and do not represent the official position of my University or Hospital Trust.

• Patient Safety: This video does not establish a doctor-patient relationship. Always seek the advice of a qualified healthcare provider regarding any medical condition.

Music generated by Mubert https://mubert.com/render

https://substack.com/@healthaibrief

#HealthAI #DigitalHealth #MedicalDevice #AmazonHealth #Telemedicine #ClinicalTech #HealthcareInnovation #HealthTech #SaMD #FutureOfMedicine

Stop searching for the next standalone medical AI app, the most powerful AI is already being built into the tools you use every day. We analyse the latest a16z "Top 100 Gen AI Consumer Apps" report to see what it means for the future of clinical digital health.

In this episode, we break down why the "AI-first" standalone product is failing and how the move toward "agentic" workflows will redefine hospital operations.

Link to the full report by Olivia Moore: https://a16z.com/100-gen-ai-apps-6/

Key Takeaways:

• How the a16z Gen AI report highlights the shift from "AI destinations" to "invisible AI operating environments."

• Why clinical workflow integration, not model power, is the primary driver of successful AI adoption.

• The critical difference between horizontal AI giants and specialized tools for high-stakes medical imaging and clinical data.

0:00 The Death of the Standalone AI Medical App

0:16 Reviewing the a16z GenAI Consumer Apps Report

0:37 AI as an Invisible Operating Environment

1:05 ChatGPT’s Evolution into a Super App

1:24 The "Extra Tab" Friction in Healthcare Workflows

1:42 The Rise of Agentic AI (Manus & OpenCoder)

2:08 Horizontal Giants vs Specialised Professional Tools

3:55 The Shift from AI as a "Fabric" Rather Than a Feature

4:26 Moving Toward "Operational Intelligence" in Health

Also catch our previous episodes on:

- Big Tech Trends in Health 2026: https://youtu.be/01fl9HMcrcc

- Agentic AI in Healthcare: https://youtu.be/eIKZ67ggW3s

- More on AI agents for workplace: https://youtu.be/5aHIBl4hNSA

- Sleep foundation model: https://youtu.be/5yvxGYtt9Vg

- TRICORDER study highlighting importance of implementation and integration within workflows: https://youtu.be/eOFZvVGKSfU

Clinical Governance & Educational Disclosure

This analysis is for educational and informational purposes only. It provides a technical review of AI in healthcare and does not constitute medical advice or treatment.

• Professional Accountability: If you are a healthcare professional, ensure your use of AI complies with local Trust policies and professional standards (GMC/NMC/HCPC).

• Evidence-Based Review: These views are my own and do not represent the official position of my University or Hospital Trust.

• Patient Safety: This video does not establish a doctor-patient relationship. Always seek the advice of a qualified healthcare provider regarding any medical condition.

Music generated by Mubert https://mubert.com/render

https://substack.com/@healthaibrief

#HealthAI #DigitalHealth #ClinicalWorkflow #MedicalInnovation #HealthTech #AIinMedicine #a16z #DigitalTransformation #HealthIT #NHSInnovation

Predict 130+ diseases from one night of sleep? Learn how the SleepFM foundation model uses AI to detect dementia, heart failure, and mortality risk up to 6 years early.

SleepFM is a breakthrough multimodal sleep foundation model trained on over 585,000 hours of polysomnography (PSG) data. By leveraging a unique "Leave-One-Out" contrastive learning approach, this AI integrates brainwaves, heart activity, and respiratory signals to create a latent representation of human health. Unlike previous supervised models, SleepFM generalizes across different clinical settings and can accurately predict the risk of conditions like Parkinson's, stroke, and chronic kidney disease years before symptoms appear.

Link to paper: https://www.nature.com/articles/s41591-025-04133-4

"A multimodal sleep foundation model for disease prediction"

Key Takeaways:

• Foundation Model for Sleep: How SleepFM uses self-supervised learning to overcome the lack of expert-labeled sleep data.

• Disease Prediction Power: Analysis of the C-Index scores for 130 conditions, including an 0.85 for dementia and 0.84 for all-cause mortality.

• Clinical Generalization: Why the "channel-agnostic" architecture allows this AI to work across different hospitals and PSG equipment configurations.

0:00 Introduction

0:27 SleepFM Overview

1:24 Technical Architecture

3:22 Disease Prediction

4:21 C-Index Definition

5:29 Model Validation

6:00 Generalization Testing

7:27 Clinical Challenges

8:40 Future Outlook

Sleep AI, Foundation Models in Healthcare, Disease Prediction, Polysomnography AI, Machine Learning in Medicine, SleepFM, Medical AI Research, Digital Biomarkers, Preventative Health AI, Neurodegeneration Detection #HealthAI #SleepMedicine #DementiaPrevention #MachineLearning #DigitalHealth #MedTech #aiinmedicine Music generated by Mubert https://mubert.com/render

[email protected]

If a patient has a 50-page record, can the AI see it all? We explain the "Context Window" and why it’s the biggest bottleneck in medical AI today.

#ContextWindow #LongContext #MedicalRecords #ai in medicine Music generated by Mubert https://mubert.com/render

[email protected]

Following some comments I’m pulling back the curtain on exactly how I use AI to produce this podcast, what I wouldn't automate, and why a recent, terrifying case of an autonomous AI executing a hit piece on a human potentially changes things.

This leads to considerations about the cognitive and operational realities of building an AI-assisted workflow. Moving beyond the "Am I an AI?" comments, we consider exact production, what is human, what is artificial, and why I threw my experiments with AI video avatars in the bin. We then dissect a chilling incident in the open-source software community where an autonomous AI agent, "MJ Rathbun," bypassed human oversight to publish a targeted reputational attack on a Matplotlib maintainer. This episode breaks down the technical mechanics of open-source AI autonomy, the psychological paradox of authenticity, and why using AI as a "cognitive forklift" is essential, but outsourcing your thinking is incredibly dangerous.

Link to the original information about he AI Agent 'hit job', including all the very thoughtful points made by Scott Shambuagh on the broader implications that I summarise in the audio, I'd highly recommend reading the original blogpost in full though: https://theshamblog.com/an-ai-agent-published-a-hit-piece-on-me/

Key Takeaways:

• The Authenticity Paradox: Why we accept physical artificial enhancement (lighting, makeup) but instinctively reject AI video avatars and voice clones due to the crucial element of deception.

• The Cognitive Forklift: Learn my exact workflow for deploying AI for efficiency (script tightening, thumbnail generation) without falling into the "Chinese Room" trap that atrophies critical thinking.

• The Threat of Autonomous Retaliation: Discover the real-world implications of decentralized AI agents (like OpenClaw) executing independent smear campaigns, and how to safeguard against misaligned autonomous logic.

00:00 - Am I an AI? Addressing questions

01:03 - Pulling back the curtain: My AI-assisted workflow

01:45 - Why the "heart" of the content must remain human

02:30 - Where AI is useful (packaging and editing)

04:32 - Failed experiments: AI video avatars and voice cloning

05:40 - The philosophy of authenticity: Human vs artificial content

06:15 - Use cases for Google’s NotebookLM

07:40 - AI avatars vs digital "enhancements" (the makeup analogy)

09:00 - Maintaining anonymity: Accountability and critique

11:25 - The "Chinese Room" and the risk of offloading thought

12:14 - Taking a forklift to the gym: The purpose of effort

13:10 - The danger of "regression to the mean" in AI models

13:42 - A parallel with Support Vector Machines (SVMs)

14:50 - Automating execution vs. automating intent

16:30 - When AI agents get personal: The MJ Rathbun story

17:15 - The surge of low-quality code contributions in open source

18:20 - An autonomous character assassination: AI vs Scott Shambaugh

19:30 - Quoting the AI’s "hit piece" on gatekeeping

20:55 - The terror of autonomous influence operations

23:01 - The "no kill switch" problem with open-source agents

24:00 - Tools that possess agency: A new threshold for technology

25:20 - Final thoughts: AI as a cognitive forklift, not a replacement

Autonomous AI Agents, Open Source AI Security, Healthcare AI Workflows, LLM Hallucinations, AI Reputational Risk, Medical AI Integration, Generative AI Authenticity, DeepMind AI Strategy, HealthTech Innovation, AI Agent Frameworks. #HealthAI #AutonomousAgents #GenerativeAI #HealthTech #CyberSecurity #aiinmedicine Music generated by Mubert https://mubert.com/render

https://substack.com/@healthaibrief

[email protected]

You don't always need a supercomputer. Learn how Quantization "compresses" AI models so they can run locally on hospital hardware.

#EdgeComputing #Privacy #LocalAI #ai in medicine Music generated by Mubert https://mubert.com/render

[email protected]

Discover how Nyra Health is using proprietary speech AI and an MDR Class IIa platform to bridge the "rehab cliff" for stroke and dementia patients.

In this deep dive, we analyse the strategic expansion of Nyra Health, a Vienna-based startup that recently secured €20M in Series A funding. We examine their "myReha" ecosystem, which uses advanced Natural Language Processing (NLP) to provide 35,000+ personalised exercises for neurological recovery. From reimbursement hurdles in the DACH region to the technical nuances of their speech models and the current state of their clinical evidence, we break down what this means for the future of digital therapeutics and neurology.

Key Takeaways:

• The Rehab Cliff: How AI platforms might maintain therapy intensity after hospital discharge.

• Clinical Integration: The role of MDR Class IIa certification and insurance reimbursement in scaling Health AI.

• Data-Driven Neurology: Utilising speech biomarkers and adaptive loops to personalise neuro-recovery.

0:00 - Introduction & €20M Series A Funding

0:12 - The "Rehab Cliff" in Neurological Care

0:54 - Nyra Health’s AI Ecosystem: myReha & nyra insights

1:41 - Clinical Engineering: Speech Models & Adaptive Feedback

2:19 - Regulatory Status & Commercial Traction

2:54 - Analyzing the Evidence: RCTs & Research Gaps

4:18 - Enhancing Therapist Efficiency with Data

4:44 - Challenges: Digital Exclusion & Displaced Care

5:19 - US Expansion & Pharma Partnerships

5:44 - The Future of Continuous Neurology

6:10 - Final Verdict & Key Takeaways

Digital Therapeutics, Neurorehabilitation AI, Stroke Recovery Tech, Nyra Health, Medical AI Reimbursement, Speech Biomarkers, Aphasia Therapy AI, HealthTech Series A, DACH Digital Health, Clinical AI Strategy, #HealthAI #Neurotech #DigitalTherapeutics #MedTech #StrokeRecovery #aiinmedicine Music generated by Mubert https://mubert.com/render

https://substack.com/@healthaibrief

[email protected]

7B, 70B, 175B - what do these numbers mean? We discuss the trade-off between LLM size, cost, and clinical accuracy.

#MachineLearning #Parameters #Llama3 #ai in medicine Music generated by Mubert https://mubert.com/render

[email protected]

Discover how AWS is using agentic AI to automate patient scheduling, documentation, and medical coding directly within the EHR.

Amazon Connect Health is a purpose-built AI solution designed to tackle the administrative complexity of modern healthcare. By integrating directly with EHRs via a unified SDK, it enables 24/7 patient verification, natural language appointment booking, and ambient clinical documentation. This system doesn't just transcribe; it uses "Evidence Mapping" to link every AI-generated note and billing code to its original source, ensuring clinical trust and auditability.

Key Takeaways

• Agentic Automation: How AI now performs real-time EHR tasks like insurance checks and scheduling without human intervention.

• Clinician Efficiency: Details on ambient documentation and medical coding tools that reduce "pajama time" and accelerate the revenue cycle.

• Trust & Verification: The technical role of Evidence Mapping in linking AI outputs to source data for clinical safety.

00:00 Introduction: Amazon Connect Health Launch

00:22 AWS as the "Connective Tissue" for EHRs

00:33 The Technical Architecture: Agentic AI Explained

01:49 Pillar 1: Streamlining Patient Engagement

02:57 Pillar 2: Point of Care (Insights & Documentation)

03:46 Accelerating the Revenue Cycle with AI Coding

04:07 Solving the "Black Box" with Evidence Mapping

05:02 AWS vs. ChatGPT: The Vertical Integration Advantage

05:34 What This Is (and Is Not): The Admin Assistant

06:21 The Future of "Invisible AI" in the Clinic

07:05 Final Verdict: Why the Sum is Greater than the Parts

Amazon Connect Health, Healthcare AI, AWS HealthLake, Ambient Clinical Documentation, Medical Coding AI, Patient Engagement AI, EHR Integration, Agentic AI Healthcare, FHIR Data, Clinical Workflow Automation. #HealthAI #AWSHealthcare #MedTech #ClinicalWorkflow #DigitalHealth #aiinmedicine Music generated by Mubert https://mubert.com/render

https://substack.com/@healthaibrief

[email protected]