AI in Medicine - curated summaries making complex issues easy to understand

We unpack the Stanford AI Index 2026 Medicine chapter to explore why AI is acing medical exams but still struggling to translate into prospective clinical trial evidence. We dive into the massive gap between FDA clearances and randomized trials, highlight which tools like ambient scribes are genuinely scaling, and reveal why smaller models are outperforming massive systems in molecular biology.

Key points

• Leading multi-agent AI systems score 85.5% on complex medical cases compared to 20% for physicians, yet nearly half of clinical AI studies still rely on simulated scenarios rather than real patient data.
• Ambient AI scribes are driving real clinical impact, with some health systems reporting an 83% reduction in note-writing effort and up to a 112% return on investment.
• The FDA authorized 258 AI medical devices in 2025, but a peer-reviewed analysis found that only 2.4% of devices with clinical studies were supported by randomized controlled trial data.
• In molecular biology, smaller specialized models are outperforming massive general-purpose systems, and biological model development is increasingly bottlenecked by data curation rather than architecture.
• AI-generated summaries now appear at the top of 84% to 92% of health-related Google searches, fundamentally shaping how patients access health information before ever seeing a clinician.

Source: Medicine - Stanford Institute for Human-Centered Artificial Intelligence, 2026

This spot is available. Reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsor the show

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/exam-scores-vs-reality-the-2026-ai-index-report-on-medicine-ec9148

We unpack the FDA 510(k) summary for Bunkerhill Contrast CAC, an AI tool cleared to opportunistically detect and quantify coronary artery calcium on routine contrast-enhanced chest CTs. We discuss the shift from non-contrast to contrast imaging, the pivotal performance data, and what it means for clinical workflows.

Key points

• Cleared for adults aged 30 years and above to detect and estimate the quantity of coronary artery calcification.
• Specifically designed to analyze non-gated, contrast-enhanced chest CT images, unlike its predicate device which uses non-contrast scans.
• Outputs an estimated volume score in mm3 rather than traditional Agatston units.
• Pivotal testing demonstrated a sensitivity of 0.957 and specificity of 0.929.
• A 50-case bridging study comparing contrast to non-contrast images showed a mean bias of -4.36 mm3.
• The software is intended as adjunctive information only and does not replace the physician's evaluation or the original patient report.

Source: FDA 510(k) summary K260166: Bunkerhill Contrast CAC (BunkerHill Health) - U.S. Food and Drug Administration, 2026

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/company-spotlight-bunkerhill-contrast-cac-opportunistic-screening-on-r-dfb2f3

The FDA is asking the healthcare industry how to regulate generative AI medical devices. We unpack their proposed competency-based approach, which evaluates AI less like software and more like a human medical resident.

Key points

• Generative AI requires new regulatory frameworks because it produces variable outputs to similar inputs and can evolve after deployment.
• The FDA proposes a two-axis framework to assess risk based on device activity (how independently it acts) and the consequences of an incorrect output.
• A new competency-based premarket evaluation would mirror human clinician credentialing, using robust benchmarking and clinical confirmation.
• Postmarket monitoring could involve periodic re-benchmarking and shared responsibility across the healthcare ecosystem to catch performance degradation.
• The FDA suggests voluntary master files for foundation models, allowing developers to share underlying safety data without making it public.

Source: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback - U.S. Food and Drug Administration, 2026

This spot is available. Reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsor the show

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/credentialing-the-ai-doctor-the-fda-s-blueprint-for-generative-ai-944c85

We unpack the FDA 510(k) summary for Tyto Insights for Wheeze Detection (with PCCP). Listeners will learn how this over-the-counter AI software analyzes lung sounds, the real performance numbers behind it, and exactly how the FDA's Predetermined Change Control Plan allows algorithms to be retrained post-clearance without new submissions.

Key points

• The software is an over-the-counter AI tool that analyzes 6 to 12-second acoustic lung signals from a compatible stethoscope to detect suspected wheezing in patients 2 years and older.
• Performance testing showed a specificity of 0.99 and a sensitivity of 0.55, meaning it rarely flags false positives but misses a significant portion of actual wheezes.
• This 510(k) clearance is based on substantial equivalence to a 2023 version of the same product, with the changes being cybersecurity updates and the addition of a Predetermined Change Control Plan (PCCP).
• The PCCP allows the company to retrain the machine learning model to improve performance without filing a new 510(k), provided the new version's sensitivity and specificity do not drop below the currently validated thresholds.
• The FDA approval letter emphasizes that even with a PCCP, all software modifications must still follow strict Quality Management System Regulations, including design controls and documentation in the Medical Device File.

Source: FDA 510(k) summary K252844: Tyto Insights for Wheeze Detection (with PCCP) (Tyto Care) - U.S. Food and Drug Administration, 2026

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/company-spotlight-tyto-care-tyto-insights-the-fda-s-blueprint-for-upda-ac2251

The AMA just surveyed nearly 1,700 physicians on their real-world use of augmented intelligence. We unpack why professional adoption has doubled, the massive hidden wave of patient AI use, and the specific liability fears keeping doctors awake at night. Essential listening for clinical leaders and medtech product teams.

Key points

• Over 80% of physicians now use AI professionally, averaging 2.3 use cases, with research summarization leading the pack at nearly 40%.
• While 76% see AI as an advantage for patient care, 88% are concerned about clinical skill loss, particularly among medical students and residents.
• There is a massive disclosure gap: 30% of doctors assume most of their patients use AI for health questions, but 29% haven't had a single patient disclose it.
• "Clear liability frameworks" ranks as the number one regulatory priority, with 88% demanding validation of safety and efficacy before adoption.
• Despite 92% of physicians wanting more AI training, surprisingly, physicians under 35 report lower training rates (55%) than those over 55 (71%).

Source: 2026 Physician Survey on Augmented Intelligence - American Medical Association, 2026

This spot is available. Reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsor the show

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/the-ai-exam-room-what-1-692-physicians-actually-want-and-fear-in-2026-35217f

We unpack the FDA 510(k) summary for Eko Health's EFAST software, which brings transformer-based AI to the digital stethoscope. Listeners will learn how the cloud-based system analyzes heart sounds and ECGs, its real-world performance numbers for detecting structural murmurs and atrial fibrillation, and the crucial caveats regarding pediatric screening and retrospective testing.

Key points

• EFAST is a cloud-based Software as a Medical Device (SaMD) that analyzes 15-second heart sound and ECG recordings via an API, returning a JSON format data structure rather than using a primary display.
• For structural murmur classification, EFAST showed an overall sensitivity of 83.4% and specificity of 86.0%, closely matching its predicate device.
• Specificity for pediatric patients dropped to 76.0%, which the document notes is likely due to confusion with innocent murmurs.
• For atrial fibrillation detection in adults, sensitivity leaped to 94.7% (compared to 73.1% in the older reference device), with a specificity of 94.1%.
• Testing was retrospective, and experts excluded 259 out of 2,460 murmur recordings and 97 out of 1,256 ECG recordings because they could not come to an agreement.
• Both test datasets heavily skewed White (90.0% for murmurs, 86.6% for ECG), which the FDA summary describes as representative of the intended use population.

Source: FDA 510(k) summary K251494: Eko Foundation Analysis Software with Transformers (EFAST) (Eko Health) - U.S. Food and Drug Administration, 2025

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/company-spotlight-eko-health-efast-transformers-hit-the-digital-stetho-989647

We dive into the FDA's recent clearance of AiTiALVSD, an AI tool that predicts left ventricular systolic dysfunction from a standard 12-lead ECG. We break down the clinical validation numbers, the demographic gaps in the training data, and what this specific 510(k) clearance means for cardiology workflows.

Key points

• AiTiALVSD by Medical AI Co., Ltd. received FDA Class II clearance to support the diagnosis of Left Ventricular Systolic Dysfunction (LVEF less than or equal to 40%) using 12-lead ECG data.
• The deep learning model was trained on 400,339 ECGs from 148,624 patients across 16 hospitals in Korea, and evaluated in a US retrospective study of 22,101 patients.
• In the US performance study, the software demonstrated a sensitivity of 83.7 and a negative predictive value of 97.7, making it a strong potential rule-out tool for clinicians.
• The FDA document explicitly notes a limitation regarding Black or African American subjects, who made up 9.6% of the validation study, noting that specificity was lower in this subgroup.
• The software provides a binary output of High or Low, but the FDA strictly warns it must not be used as a substitute for echocardiography or as a standalone diagnosis.

Source: Re: K254302 - U.S. Food and Drug Administration, 2026

This spot is available. Reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsor the show

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/fda-clears-ai-for-heart-failure-unpacking-the-aitialvsd-510-k-8822ad

We unpack a critical Stanford and UC Berkeley report advising California on its new Transparency in Frontier AI Act (TFAIA). Discover why current legal definitions of AI risk are dangerously narrow, how a strict focus on computing power ignores biological threats, and what new reporting thresholds could mean for healthcare and medtech developers.

Key points

• A computing threshold of 10^26 operations is an imperfect proxy for risk, as smaller models trained on biological data can still pose catastrophic threats.
• The original law defines catastrophic risk as 50 deaths or 1,000,000,000 dollars in property damage, which fails to capture cumulative harms like mental health deterioration.
• Basing compliance on 500,000,000 dollars in revenue misses highly funded startups; the report suggests adding a 100,000,000 dollars R&D spending threshold.
• Internal testing incidents, such as loss of control within a developer's sandbox, should be reportable even if they do not cause immediate physical harm.
• Meaningful AI evaluation requires a sociotechnical approach that looks beyond the model to the tools, workflows, and human interactions surrounding it.

Source: Implementing the Transparency in Frontier AI Act (TFAIA) Report for the California Department of Technology Multistakeholder Workshops - Stanford Institute for Human-Centered AI

This spot is available. Reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsor the show

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/when-ai-escapes-the-lab-redefining-catastrophe-in-california-s-frontie-d9e2fd

NICE weighs in on 4 AI technologies designed to speed up echocardiography for heart failure diagnosis. Despite incredible claims of time savings, the agency stops short of recommending NHS funding. Maya and Sam unpack the evidence gaps, the economic models, and what AI builders need to prove before winning health-system adoption.

Key points

• NICE published an early-use assessment in May 2026 on 4 AI echocardiography tools: EchoConfidence, EchoGo Heart Failure, Ligence Heart, and Us2.ai.
• None of the technologies are recommended for NHS funding yet; access should remain through company, research, or non-core funding.
• While AI reduced analysis time in some studies—from 587 seconds to 3.2 seconds—it is uncertain if this translates to reduced waiting lists in the real world.
• The clinical evidence lacked generalisability, with many studies being retrospective, outside the UK, and excluding poor-quality scan images.
• Economic modeling showed potential cost savings or acceptable cost-effectiveness, but results are highly uncertain without real-world procedure time data.

Source: Artificial intelligence (AI)-assisted echocardiography analysis and reporting to support the diagnosis and monitoring of heart failure: early-use assessment - National Institute for Health and Care Excellence

This spot is available. Reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsor the show

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/can-ai-clear-the-echocardiogram-backlog-nice-s-verdict-on-heart-failur-95f0af

We unpack the FDA 510(k) summary for icobrain aria by icometrix, an AI tool cleared to help radiologists detect brain swelling and bleeding in Alzheimer's patients. We explore the clinical trial data, the significant jumps in sensitivity for spotting mild abnormalities, and what it means for the rollout of new amyloid beta-directed therapies.

Key points

• icobrain aria is a software-only device that assists radiologists in detecting ARIA (amyloid-related imaging abnormalities) from MRI scans of Alzheimer's patients.
• In a clinical study with 16 radiologists, the software increased sensitivity for detecting mild ARIA-E from 47.2% to 70.2%.
• The software provides color-coded 3D overlays and a two-page electronic report, tracking changes in edema and microhemorrhages across two time points.
• It does not replace clinical judgment and cannot segment macrohemorrhages of 10 mm or more.
• FDA clearance was based on substantial equivalence to OsteoDetect, highlighting how AI in radiology is evaluated on workflow and algorithmic similarities rather than identical disease states.

Source: FDA 510(k) summary K240712: icobrain aria (icometrix NV) - U.S. Food and Drug Administration, 2024

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Full transcript: https://ai-in-medicine-podcast.vercel.app/episodes/company-spotlight-icometrix-icobrain-aria-the-ai-unlocking-alzheimer-s-fcc19a