ChatGPT Health. ChatGPT for Clinicians. ChatGPT for Healthcare. MedGemma. Dragon Copilot. Claude for Healthcare. In 2026, every major tech company released a healthcare AI product — and most of them sound like variations of the same thing. This guide explains what each one actually does and who it is for.

⏱️ Reading time: ~4 minutes

🔐 One thing to know first: data and the law

Several US tools are not available in Europe. OpenEvidence — a popular clinical AI used by hundreds of thousands of verified physicians in the US — pulled out of the EU and UK in early 2026, citing regulatory uncertainty [1]. This matters for any US tool you consider: US law includes provisions (the CLOUD Act) that can require American companies to share stored data with US authorities, regardless of where that data is held — a conflict with GDPR (General Data Protection Regulation) that remains legally contested.

🧠 OpenAI: three products, one confusing name

OpenAI has three healthcare products. They all say “ChatGPT.” They are not the same.

• ChatGPT Health — for patients. Explains lab results in plain language, syncs with fitness trackers.

• ChatGPT for Clinicians — for verified doctors. Clinical questions and evidence search with peer-reviewed citations.

• ChatGPT for Healthcare — for hospitals. A version of ChatGPT configured to know your hospital’s own guidelines, drug lists, and clinical workflows. Handles documentation, patient intake, and administrative tasks.

The one most physicians ask about is ChatGPT for Clinicians. It is free and includes a research mode that cites peer-reviewed sources [2]. According to OpenAI, hundreds of medical advisors reviewed over 700,000 model responses — with 99.6% rated as safe and correct in clinical testing [2]. Notably, the tool showed no advantage over physicians in direct patient consultation tasks [2]. The catch for European physicians: verification requires an NPI (National Provider Identifier), a US-specific credential. A rollout across other countries is planned but not yet confirmed [2] — For readers who want to explore the Swiss context in more detail, this Medinside article provides additional background and perspective (German) Medinside article (DE).

🔬 Google MedGemma 1.5: powerful imaging AI — but not for doctors yet

MedGemma 1.5 was released in January 2026 [3]. It can analyse 3D CT and MRI scans, tissue samples under the microscope, and chest X-rays — significantly better than its predecessor [3]. It is not an app you open: it is open-source code that software developers use to build new tools. Google is explicit that it is not validated for direct clinical use [3]. Think of it as the engine inside future radiology support tools — not something to use in practice today.

🛠️ Microsoft Dragon Copilot: ambient documentation

Dragon Copilot listens to your patient consultation and automatically writes a structured clinical note, which you then review and sign. In 2026, Microsoft expanded it beyond note-writing: it now suggests diagnostic codes using ICD (International Classification of Diseases) terminology and can draft referral letters in real time [4]. It connects to major EHR (Electronic Health Record) systems including Epic and is rolling out across Europe in 2026 [4].

🧾 Anthropic Claude for Healthcare: not really for physicians

Claude for Healthcare is built for hospital billing teams and insurance authorisation workflows. It connects to ICD coding registries and is designed for administrative and health-tech teams — not individual clinicians.

🇨🇭 Meditron: the Swiss option

Meditron is an open-source medical AI model developed at EPFL (Swiss Federal Institute of Technology Lausanne), trained on peer-reviewed medical literature and international clinical guidelines [5]. Its main advantage: it runs entirely within a hospital’s own servers, so patient data never leaves the building. Following an evaluation with healthcare staff using hypothetical clinical cases in 2025, the CHUV (Centre Hospitalier Universitaire Vaudois) is set to begin testing Meditron in their emergency department in May 2026 [5].

Beyond the US AI: Made in China

Worth knowing: DeepSeek-R1, a Chinese open-source model, performs comparably to proprietary AI on clinical benchmarks according to a Nature Medicine study [6] — and like Meditron, can be run locally. However, its Chinese origin raises data sovereignty questions for European institutions.

👤 For patients: ChatGPT Health and Perplexity Health

Two consumer-facing tools are worth knowing about when patients ask. ChatGPT Health (OpenAI) syncs with fitness trackers and explains lab results in plain language.
Perplexity Health, launched in March 2026, connects Apple Health, wearables, and health records from over 1.7 million care providers to answer personal health questions [7]. Both are designed to help patients prepare for a doctor’s visit — not to replace one.
However, early independent evidence suggests caution: recent studies show that general-purpose medical AI can miss urgent scenarios and should not be relied on without clinical oversight (see No.4: AI in Research: 5 Recent AI Studies).

🔎 What each tool does — at a glance

ChatGPT for Clinicians — clinical questions and evidence search with citations. For physicians (US only for now).

Dragon Copilot — writes your clinical notes during the consultation. For physicians.

MedGemma 1.5 — interprets CT, MRI, and histopathology. Foundation for future tools; not yet for direct clinical use.

Meditron — Swiss open-source AI that runs locally on hospital servers. For institutions.

Claude for Healthcare — medical coding and insurance workflows. For admin and billing teams.

ChatGPT Health / Perplexity Health — personal health questions using wearable and health record data. For patients.

Share

Subscribe now

📚 Sources

[1] OpenEvidence EU/UK withdrawal — Let’s Data Science

[2] ChatGPT for Clinicians — Medinside.ch (DE); HIT Consultant

[3] MedGemma 1.5 — Google Research Blog

[4] Dragon Copilot — ORdigiNAL; HIT Consultant

[5] Meditron / CHUV — SWI swissinfo.ch

[6] DeepSeek-R1 clinical benchmarks — Nature Medicine

[7] Perplexity Health — Modern Healthcare

Ambient AI scribes are tools that listen to the clinical conversation in real time and automatically generate a draft note — no typing required. Their adoption in clinical settings is accelerating rapidly. They have a strong case for reducing physician burnout and documentation time [1,2]. What’s less discussed is what they change on the other side of the room – the patient side, and not always for the better.

⏱️ Reading time: ~3 minutes

🔧 What Scribes Actually Deliver

The evidence is real but more modest than the commercial narrative suggests:

• The largest RCT to date (238 outpatient physicians, 14 specialties) found one of two scribes tested (Microsoft Dragon Ambient eXperience (DAX) Copilot or Nabla) reduced documentation time by ~10%; the other showed no significant effect [1]

• Both reduced burnout-related metrics

• A pre-post study (46 clinicians, 5 weeks) found documentation time dropped from 10.3 to 8.2 minutes per appointment, and after-hours documentation reduced by 30% — but without a control group [2]

• One study found only 34 seconds saved per note on average — with significant variability between physicians [6]

• Note accuracy: “correct 80–85% of the time” for subjective sections — meaning edits are still needed before every sign-off [2]. AI hallucination rates are estimated at 1–3% [6]

• Consistently reported benefit: less divided attention, more eye contact with patients [2]

What hasn’t been shown yet: any improvement in patient outcomes.

🧠 What Patients Know (At the moment, not much)

• Fewer than 1 in 3 patients are aware AI scribes already exist in clinical settings [3]

• When informed, ~60% express reluctance

• 74.9% report initial comfort — but this drops to 55.3% once they’re told about AI features, data storage, and corporate involvement [4]

• Privacy concerns are the strongest predictor of rejection

In practice, consent often means a notice in the waiting room or a brief verbal mention. Whether patients understand that the entire conversation is being transcribed is a different question.

🚨 The Self-Censorship Problem

A 2025 study found that when patients know an AI is recording, a significant share would withhold information on [4]:

• Mental health — 35%

• Sexual health — 40.8%

• Illicit activity — 51.5%

The clinical risk is direct. Ambient scribes are adopted partly to free up physician attention for the patient. But if patients self-censor on exactly the topics that require openness, the consultation becomes more documented and less informative at the same time.

💡 Key Takeaways for Your Practice

1. A notice on the door is not enough A brief verbal opt-in at the start — “there’s a tool recording today to help with my notes — is that okay?” is recommended. Ideally, written consent should also be obtained, for example as part of the data privacy statement required at the start of treatment.

2. Don’t wait for patients to volunteer sensitive information When something is listening, patients are less likely to bring up mental health, sexual health, or substance use unprompted. Ask directly.

3. Reconsider ambient recording for sensitive consultations For psychiatry, addiction, sexual health — the trade-off between documentation efficiency and patient disclosure deserves a deliberate decision, not a default.

4. Language matters Tools trained predominantly on English perform less reliably on other languages [5]. So in the Swiss context, where the patient population is linguistically diverse, accuracy is not a given.

5. Marginalised patients may pull back more Patients who already approach healthcare with caution — migrants, asylum seekers, undocumented patients — may self-censor more than average. Worth keeping in mind before applying ambient recording across the board.

❓ What We Don’t Know Yet

The evidence on ambient scribes is still measuring the wrong endpoint. Reduced documentation time matters — but it’s a proxy. No study has yet linked ambient scribes to better patient outcomes. And the documentation gap runs deeper than self-censorship: one analysis found that approximately 50% of patient problems discussed verbally were never captured in the EHR at all [6] — regardless of whether AI was involved. The self-censorship data is largely self-reported and comes from small samples. We also don’t know whether patients become more comfortable over time as the technology becomes familiar, or whether the withholding effect persists — and quietly compounds.

Subscribe now

Share

📚 Sources

1. Lukac PJ, Turner W, Vangala S, et al. Ambient AI Scribes in Clinical Practice: A Randomized Trial. NEJM AI. 2025;2(12). https://pmc.ncbi.nlm.nih.gov/articles/pmid/41497288/

2. Duggan MJ, Gervase J, Schoenbaum A, et al. Clinician Experiences With Ambient Scribe Technology to Assist With Clinical Documentation. JAMA Netw Open. 2025;8(2):e2460637.https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2830383

3. Chandrasekaran R, Moustakas E. Patient Attitudes Toward Ambient Artificial Intelligence Scribes in Clinical Care.J Am Med Inform Assoc. 2026;33(2):263–272. https://doi.org/10.1093/jamia/ocaf218

4. Lawrence K, Kuram VS, Levine DL, et al. Informed Consent for Ambient Documentation Using Generative AI in Ambulatory Care. JAMA Netw Open. 2025;8(7):e2522400. https://doi.org/10.1001/jamanetworkopen.2025.22400

5. Kakani P, Kilaru AS, Buntin MB. Implications of Artificial Intelligence–Powered Ambient Scribes. JAMA Health Forum. 2026;7(1):e256150. https://doi.org/10.1001/jamahealthforum.2025.6150

6. Topaz M, Peltonen LM, Zhang Z. Beyond human ears: navigating the uncharted risks of AI scribes in clinical practice. NPJ Digit Med. 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12460601/

Disclosure: No conflict of interest. Sources independently identified by SwissMedAI from peer-reviewed literature and published editorials.

Healthcare AI in Switzerland is moving fast across institutions, but until now there has been no single place to see where, by whom, and with which tools.

We have systematically mapped where artificial intelligence is being deployed across the Swiss healthcare system. Today, we are making that work publicly available.

→ mdai.ch/ai-map

This map is designed to give clinicians, healthcare leaders, policymakers, and developers a clearer, more structured overview of how AI is already being used across Swiss healthcare.

🔎 What you can explore

The Swiss Healthcare AI Map documents where AI tools are being used across Swiss hospitals, clinics, and outpatient practices — and makes those deployments searchable by region, setting, specialty, and use case.

So far, the map includes:

• 150 documented deployments

• 38 AI tools tracked

• 22 cantons represented

📚 Methodology

Inclusion in the map requires traceable public evidence, such as a press release, institutional case study, vendor reference, or named clinical testimonial, linking a Swiss institution to a specific AI solution.

Sources were compiled through desk research, vendor reference lists, and direct submissions from institutions and developers.

📌 Scope and limitations

Due to the decentralised structure of Swiss healthcare and the fact that many deployments are not publicly communicated, actual AI adoption is likely broader than what appears here.

The map should be understood as a structured overview rather than a definitive census or independent audit.

📩 Contribute

This map improves with every contribution.

If your institution is using an AI tool not yet included, or if you notice an inaccuracy, we welcome your input.

→ Submit a deployment

Share

Subscribe now

Reading time: ~3 minutes

🩺 Can AI Actually Improve a Doctor’s Diagnosis?

Qazi et al. | Nature Health | 2026 | Large language model diagnostic assistance for physicians in a lower-middle-income country: a randomized controlled trial

Study design: Randomized controlled trial

Participants: 60 physicians recruited, 58 completed in Pakistan

Task: Six expert-developed diagnostic cases

This trial tested whether GPT-4o could improve physicians’ diagnostic reasoning. Doctors with AI support performed clearly better than those using standard resources alone, without a major time penalty.

Why it matters

This is one of the strongest recent studies showing that AI can support clinical thinking, not only documentation.

Key takeaway

AI may improve diagnostic reasoning, but the challenge is to use that support without becoming too dependent on it.

🔄 What Happens When AI Shapes the Referral Before the Specialist Visit?

Tao et al. | Nature Medicine | 2026 | An LLM chatbot to facilitate primary-to-specialist care transitions: a randomized controlled trial

Study design: Randomized controlled trial

Participants: 2,069 patients, 111 specialists, 24 disciplines

Task: AI-assisted preassessment before specialist consultation

This chatbot collected history, suggested preliminary diagnoses, proposed tests, and generated referral reports before the patient saw the specialist. The study found shorter consultation times and better physician-rated coordination and patient communication.

Why it matters

This is not just an efficiency tool. It shows that AI is starting to influence how cases are framed before specialist evaluation begins.

Key takeaway

AI does not need to make the final diagnosis to shape care. If it structures the referral, it may also influence later clinical reasoning.

⚠️ Can AI Advice Quietly Make Clinical Decisions Worse?

Kucking et al. | International Journal of Medical Informatics | 2026 | Impact of AI recommendation correctness on diagnostic accuracy in clinical decision-making

Study design: Simulated diagnostic intervention study

Participants: 223 physicians and nurses

Task: 1,338 diagnostic decisions with correct or incorrect AI recommendations

This study tested what happens when clinicians receive good or bad AI advice. When the AI recommendation was correct, performance improved. When it was incorrect, clinicians performed worse.

Why it matters

This is one of the clearest examples of automation bias in medicine. The danger is not only that AI can be wrong, but that clinicians may trust it when it is wrong.

Key takeaway

One of the main safety problems in medical AI is the combination of model error and human trust.

🚨 How Safe Is AI When It Has to Recognize an Emergency?

Bickmore et al. | Nature Medicine | 2026 | ChatGPT Health performance in a structured test of triage recommendations

Study design: Structured vignette-based evaluation

Cases: 60 clinician-authored vignettes across 21 clinical domains under 16 conditions

Output: 960 responses

This study evaluated ChatGPT Health in urgent care triage. The findings were concerning. Among gold-standard emergency cases, the system under-triaged 52%. The paper specifically describes dangerous under-triage in diabetic ketoacidosis and impending respiratory failure, including advice for review in 24 to 48 hours instead of emergency care.

Why it matters

This is a clear reminder that a system may sound medically competent while still giving unsafe advice in time-sensitive situations.

Key takeaway

Good communication does not equal safe clinical judgment. In triage, that gap can be dangerous.

🧠 Is AI Becoming More Than Just a Clinical Assistant?

Goh et al. | Nature Medicine | 2025 | GPT-4 assistance for improvement of physician performance on patient care tasks: a randomized controlled trial

Study design: Prospective randomized controlled trial

Participants: 92 practising physicians

Task: Five expert-developed clinical vignettes based on real de-identified patient encounters

This trial focused on patient care and management reasoning rather than diagnosis alone. Physicians using GPT-4 together with standard resources performed better than controls.

Why it matters

This is an important physician-focused trial because it shows that AI may support broader clinical reasoning, not just narrow question answering.

Key takeaway

AI is increasingly acting as a clinical copilot. The question is no longer whether it has value, but how to use it without weakening independent judgment.

🧾 Bottom line

These studies show that medical AI is becoming too relevant to daily practice to ignore. It may improve diagnostic reasoning, support care transitions, and assist with patient management. At the same time, its performance can vary across clinical scenarios, highlighting the need for continued development and careful human oversight, particularly in high-risk settings such as triage.

Subscribe now

📚 Sources

• Qazi et al. Nature Health (2026): Large language model diagnostic assistance for physicians in a lower-middle-income country: a randomized controlled trial

• Tao et al. Nature Medicine (2026): An LLM chatbot to facilitate primary-to-specialist care transitions: a randomized controlled trial

• Kucking et al. International Journal of Medical Informatics (2026): Impact of AI recommendation correctness on diagnostic accuracy in clinical decision-making

• Bickmore et al. Nature Medicine (2026): ChatGPT Health performance in a structured test of triage recommendations

• Goh et al. Nature Medicine (2025): GPT-4 assistance for improvement of physician performance on patient care tasks: a randomized controlled trial

AI tools in medicine are developing quickly. Many promise to save time, but it is often difficult to know which tools are truly useful and which ones fit our expectations for privacy, data governance, and clinical reality.

To make this easier, SwissMed AI is launching the AI Tools Directory: a curated overview of AI tools, organized for everyday clinical work, with a Swiss focus.

Who this is for

This directory is for physicians and medical professionals who want to:

• reduce documentation and administrative workload and work more efficiently.

• compare tools that are transparent about data residency and privacy.

• Imaging, diagnostic and decision support AI tools will be added separately in the future.

Why a dedicated directory?

The goal is not just to list tools, but to support a structured comparison:

• Clear categories by workflow: Tools are grouped by use case (e.g., Ambient Scribes, Documentation, Billing & Coding, Administration, Clinical Research).

• Focus on privacy and governance: For each tool, we summarize publicly available information relevant to Switzerland’s revised Federal Act on Data Protection (FADP; nDSG) and, where applicable, GDPR (General Data Protection Regulation) considerations, including data residency (Switzerland/EU), subprocessors, retention, and whether inputs may be used for model training.

• Swiss context: Preference is given to tools that support the linguistic and regulatory needs of Swiss healthcare.

A resource in progress

This directory is an initial version. New tools will be added over time, and existing entries will be updated regularly as products and policies change.

What you will find today:

• A searchable list of AI tools, categorized by clinical use case

• Initial, vendor-stated information on privacy, security, and data residency

• Direct links to explore the tools

Potential future directions:

• Peer input: Optional physician feedback, where feasible.

• Pricing overview: Pricing indicators where vendors publish them or choose to share them.

• Evidence references: Links to research or evaluations when available and applicable.

• HIN compatibility: where applicable.

How you can contribute

The most reliable insights come from real clinical use. If you are using a tool that has improved your workflow or if you have identified risks or limitations, your feedback will help keep the directory accurate and clinically relevant.

👉 Explore the SwissMed AI Tools Directory

Share

Subscribe now

Disclaimer: This directory is an informational resource based on vendor documentation and other publicly available sources. It does not provide legal advice or certify compliance, and details may change over time.

Reading time: approx. 3 minutes

🏥 Which Swiss hospitals are already using Epic?

🟢 LUKS Group – Live since 2019; Spital Nidwalden integrated in early 2024

🟢 Insel Gruppe / Inselspital Bern – Live since March 2024

🟡 USZ Zurich – Contract awarded August 2025 (CHF 94.8M); go-live earliest 2027

🟡 Kispi Zurich – Submission award November 2023 (up to CHF 51.2M); contract negotiations underway, go-live by end of 2027

🟡 CHUV + 11 Vaud hospitals – Contract awarded November 2025, following a legal dispute that reached the Federal Supreme Court

🔵 Canton of Bern (all cantonal hospitals) – Legislative revision currently in public consultation

(Sources: Medinside, Netzwoche, Inside-IT)

The bottom line: if current timelines hold, several Swiss university hospitals will be running Epic in parallel by 2027.

🤖 What Epic’s AI can already do

🎙️ Ambient documentation

The system listens during consultations, transcribes the conversation, and drafts a clinical note. You review and adjust — rather than starting from scratch. Epic’s own figures cite a meaningful reduction in documentation time; independent evidence remains limited for now.

(Sources: Epic, spsoft.com)

🧠 Decision support embedded in your workflow

While you navigate the chart, Epic runs in the background — analyzing labs, vitals, and medications — and surfaces alerts where you actually need them: sepsis warnings, readmission risk, treatment comparisons with similar patients.

The underlying dataset is called Cosmos: roughly 300 million anonymized patient records and over 18 billion clinical encounters from over 300 health systems worldwide.

One important caveat: the vast majority of this data comes from the US. How well these models handle Swiss clinical guidelines, SwissDRG coding, and our patient population remains largely unstudied.

(Source: cosmos.epic.com)

💰 Administrative support

Penny is Epic’s revenue cycle coding agent — it suggests codes, drafts insurance appeals, and automates claims follow-up. For now, it is clearly oriented toward the US market.

(Source: CNBC)

🔭 What’s coming: three new AI assistants

Epic’s annual customer conference — the Users Group Meeting (UGM) — took place in August 2025. Three new tools were announced:

Art (for clinicians) Prepares context for your next action: patient history, medications, outstanding results. Writes clinical notes directly from the conversation (AI Charting, developed in collaboration with Microsoft Dragon). Available in the US from early 2026; Swiss rollout date not yet announced. (Sources: CNBC, Healthcare IT Leaders)

Emmie (for patients) A chatbot integrated into the MyChart patient portal — explains results, suggests appointments. MyChart is less widely used in Switzerland than in the US, though some institutions like Inselspital already operate an Epic-based patient portal. In the Swiss context, Emmie remains of limited relevance for now. (Source: Healthcare IT News)

Curiosity (new prognostic model) Simulates potential disease trajectories: does this patient need admission? How long? Is there a risk of 30-day readmission? Trained on over 100 billion clinical events, developed with Yale and Microsoft. Pilot deployments are expected to begin in 2026; clinical use in Switzerland is likely still years away. (Sources: Epic, HIT Consultant)

🇨🇭 What this means in practice for Switzerland

LUKS and Inselspital receive AI updates automatically with Epic’s regular biannual releases. As early as May 2024, physicians at Inselspital could access a LUKS patient’s Epic record directly within the system — no fax, no PDF, in real time. This marked the first live, system-native data exchange between two Epic hospitals in Switzerland. (Source: luks.ch)

USZ, Kispi, and CHUV are still in implementation planning. For colleagues working at those institutions: now is the right moment to understand which AI features are on the roadmap — before IT and management make those decisions without clinical input.

On data privacy: The US Cloud Act continues to generate political debate. The law allows US authorities to request data from US companies under certain circumstances — regardless of where the servers are located. Dozens of Zurich cantonal council members have filed interpellations on the topic. USZ and LUKS both emphasize that data is stored on Swiss servers. (Sources: Medinside, Inside-IT)

Worth keeping in mind: all affected hospitals procured Epic with full awareness of this legal situation. The question is less whether to accept this residual risk, and more how to manage it — for example, which contractual safeguards Epic provides.

AI can genuinely take work off our plates — but only when the models are actually built for our context. If you work in an Epic hospital, it’s worth asking: which features are already switched on?

Do you work in an Epic hospital? I’d love to hear from the field — write to me at hello@mdai.ch or leave a comment below.

Share

Subscribe now

Sources

Medinside.ch · Netzwoche.ch · Inside-IT.ch · luks.ch · cosmos.epic.com · CNBC · Healthcare IT Leaders · Healthcare IT News · HIT Consultant · spsoft.com

This is a curated shortlist of projects where AI does something tangible: guiding a procedure, coaching an exam, automating interpretation, or running a closed loop.

Legend: 🟢 Now in use | 🟡 Early clinical | 🔴 Research

Cardiology

TAVIPILOT Soft (Caranx Medical) 🟡 Early clinical

What it is: AI software that tracks anatomical landmarks during TAVI/TAVR and gives the operator real-time visual guidance for valve positioning.

Why it matters: It aims to make critical positioning steps more consistent by translating live procedural information into clearer guidance at the moment decisions are made.

Ultreon™ 2.0 (Abbott) AI-enabled OCT for PCI optimisation 🟢 Now in use

What it is: An AI-enabled OCT intravascular imaging workflow that automatically highlights key features (e.g., calcified plaque) and supports stent planning and post-PCI assessment (e.g., expansion checks) during coronary interventions.

Why it matters: It turns high-detail OCT images into actionable measurements and flags in real time, reducing manual interpretation and supporting more standardised PCI decision-making.

Oncology

EUCAIM / Cancer Image Europe (European Cancer Imaging Initiative) 🔴 Research infrastructure

What it is: A Europe-wide, federated platform that lets hospitals contribute and access large collections of cancer imaging data so AI models can be trained, tested, and benchmarked across diverse datasets.

Why it matters: The bottleneck for reliable imaging AI is rarely the algorithm. It’s data scale + validation across sites. This infrastructure is designed to make AI tools less “single-center demo” and more generalisable.

Ultrasound

ALISSE (European Space Agency project) 🔴 Research

What it is: An AI system that guides users in real-time to capture diagnostic-quality ultrasound images, originally developed to help astronauts perform scans independently during deep-space missions where Earth-based guidance isn’t feasible. The technology is now being adapted for point-of-care and remote settings on Earth.

Why it matters: Ultrasound fails when you don’t capture the right plane. ALISSE aims to make consistent scanning accessible to less-experienced users, addressing workforce shortages in primary care and remote locations.

ARTHUR V.2.0 + DIANA (Robotic ultrasound + AI model) 🟡 Early clinical (published)

What it is: A robotic ultrasound system that performs standardised scanning, paired with an AI model that detects and grades ultrasound features of synovitis in rheumatoid arthritis.

Why it matters: Two steps are being automated: (1) scan acquisition (the robot makes it consistent) and (2) interpretation (the AI makes scoring more reproducible). The goal is less operator variability between sites and users.

Diabetes

CamDiab fully closed-loop insulin algorithm (CE marked) 🟢 Now in use

What it is: A fully closed-loop automated insulin delivery algorithm designed to adjust insulin continuously, without meal boluses or carb counting (as described in the CE-mark news report).

Why it matters: It aims to reduce the daily cognitive load of diabetes management by shifting decisions from the user to an automated control system. (Some debate whether to call this “AI/ML” versus advanced control, but the clinical impact is the point.)

Endoscopy

AICE (Horizon Europe) 🔴 Research

What it is: AI designed to analyse colon capsule endoscopy video frames and flag likely findings, reducing manual review burden.

Why it matters: Capsule endoscopy produces massive numbers of images per patient. The goal is to make review faster and scalable, so capsule screening becomes more practical outside niche use cases.

AI-adjacent “physical medicine” worth watching

ETH Zurich microrobots (robotics/micro-engineering) 🔴 Research

What it is: Magnetically guided microrobot capsules that can carry a drug payload and release it at a specific target site (described as dissolvable, targeted delivery in the ETH write-up).

Why it matters: It points toward therapy that is physically steered to the site of action instead of relying on systemic dosing. The core innovation here is microrobotics + magnetic control, not machine learning directly.

The pattern across these innovations is consistent: AI is transitioning from retrospective analysis to real-time guidance, from single-institution validation to federated infrastructure. What began as computer vision research (detecting polyps, identifying synovitis) is becoming integrated clinical tooling. The timeline varies, some tools are commercially available now, others require 3-5 years of validation, but the trajectory is clear.

Share

Subscribe now

Sources

• TAVIPILOT Soft: Cardiac Interventions Today. “Caranx Medical’s TAVIpilot Soft Used in First Clinical Cases.” November 2025. https://citoday.com/news/caranx-medicals-tavipilot-soft-used-in-first-clinical-cases | SAITO-1A trial (NCT07043075)

• Ultreon 2.0: Abbott Cardiovascular. “Ultreon™ 2.0 Software for OCT Intravascular Imaging.” https://www.cardiovascular.abbott/de/de/hcp/percutaneous-coronary-intervention/intravascular-imaging/ultreon-software/ultreon-2-0.html | PMC11993893

• EUCAIM: Cancer Image Europe. https://cancerimage.eu | European Commission. “European Cancer Imaging Initiative.” https://digital-strategy.ec.europa.eu/en/policies/cancer-imaging

• ALISSE: ESA. “Just add AI for expert astronaut ultrasound.” March 2024. https://www.esa.int/Enabling_Support/Space_Engineering_Technology/Just_add_AI_for_expert_astronaut_ultrasound | Nebula ESA. https://nebula.esa.int/content/autonomous-ultrasound-image-improvement-system-alisse

• ARTHUR V.2.0 + DIANA: Frederiksen BA, et al. “Automated ultrasound system ARTHUR V.2.0 with AI analysis DIANA V.2.0 matches expert rheumatologist in hand joint assessment of rheumatoid arthritis patients.” RMD Open. 2025. PMID: 40764087

• CamDiab: Drug Delivery Business News. “CamDiab earns nod for fully closed-loop insulin delivery algorithm.” January 2026. https://www.drugdeliverybusiness.com/camdiab-ce-mark-fully-closed-loop-insulin-algorithm/

• AICE: AICE Project. https://aiceproject.eu

• ETH Zurich microrobots: ETH Zurich. “Microrobots finding their way.” November 2025. https://ethz.ch/en/news-and-events/eth-news/news/2025/11/microrobots-finding-their-way.html

A 2026 study in Rheumatology International compared three AI tools head-to-head to test a simple question: do design, certification, or cost translate into better diagnostic performance?

The AI Tools Tested

• Prof. Valmed: Subscription-based, EU-certified medical device (RAG-based)

• ChatGPT-5 Thinking: General-purpose LLM (subscription)

• OpenEvidence: Free, RAG-based tool for healthcare professionals

How the Study Worked

Researchers entered 60 clinical vignettes (rare rheumatic diseases plus relevant differentials) into each system using the same standardized prompt. Each tool returned five diagnostic suggestions with diagnostic probabilities.

Three blinded rheumatologists rated each suggestion as:

• Identical to the correct diagnosis

• Plausible

• Diagnostically different

Performance was reported as:

• Top-1 accuracy (identical top diagnosis)

• Top-5 performance (identical or plausible diagnoses among the five)

• A total diagnostic score (2 points for identical, 1 point for plausible)

• Processing time

Key Results (in plain numbers)

Identical top diagnosis (Top-1):

• Prof. Valmed: 23.3%

• ChatGPT-5 Thinking: 26.7%

• OpenEvidence: 35.0%

Identical or plausible within top 5:

• Prof. Valmed: 51.7%

• ChatGPT-5 Thinking: 58.3%

• OpenEvidence: 56.7%

Total diagnostic score (overall quality across outputs):

• Prof. Valmed: 212

• ChatGPT-5 Thinking: 226

• OpenEvidence: 221

Processing time (mean):

• Prof. Valmed: 20 s

• ChatGPT-5 Thinking: 36 s

• OpenEvidence: 31 s

Bottom line statistically: differences between systems were not significant in pairwise comparisons.

What Stands Out

• No clear winner. All three systems performed broadly similarly on rare rheumatology vignettes.

• Certification did not translate into higher accuracy in this benchmark. The certified medical device did not outperform the others.

• AI “confidence” behaved sensibly. All systems assigned higher probabilities to correct diagnoses than to incorrect ones.

• General-purpose AI held up. ChatGPT-5 Thinking was competitive and achieved the highest overall diagnostic score.

• Speed was a non-issue. All tools delivered outputs within a minute.

Clinical Interpretation

Even the best-performing tool produced the correct top diagnosis in only 35% of cases. That is a reminder, not a punchline: these systems can help generate differentials, but they are not reliable standalone diagnostic engines for rare rheumatic disease.

Limitations

• Vignettes, not real patients: Controlled case summaries don’t reflect real-world workflow or messy data.

• Possible prior exposure: Some cases came from published sources, so data leakage can’t be fully excluded.

• One prompt only: Results may change with different prompting or interaction styles.

• Small sample: 60 cases may miss small but meaningful differences.

• Snapshot in time: Models evolve fast, so performance may not hold for future versions.

  ---

Bottom Line

No clear winner, but a clear benefit: AI can broaden your differential in seconds and surface diagnoses you might not think of in a busy shift, with clinician oversight guiding the final decision.

Subscribe now

Share

Source: Kremer P, Langballe E, Haase I, et al. Diagnostic performance of Prof. Valmed, ChatGPT-5 Thinking, and OpenEvidence in rheumatology: A comparative evaluation. Rheumatology International (2026). https://doi.org/10.1007/s00296-025-06068-y

To avoid duplicate emails: which language do you prefer for SwissMed AI emails?

Update email language

The practical answer: use a hybrid workflow

• Input: your clinical working language (DE/FR/IT)

• Process instruction: “Analyze using international medical literature and standard medical terminology.”

• Output: the language you need for documentation and handoffs

Why this works: it separates reasoning from documentation clarity and keeps your prompts auditable.

Example: 03:00 admission note in Swiss German → prompt “Summarize using standard medical terminology” → structured output in German for your clinical information system.

Subscribe now

Four key facts

1) English is the default knowledge substrate

As of December 2023, 86.5% of PubMed-indexed publications were in English.
That matters because training and evaluation data often reflect what is most available.

2) “English is always better” is not a fact

Multilingual performance varies by model and task, so claims should be benchmark-based, not anecdotal. Medical multilingual benchmarks like MedExpQA exist specifically to measure cross-language differences.
Takeaway: English is often a reasonable default for complex reasoning, but it is not a safety guarantee.

3) The bigger risk is messy input

What reliably worsens outputs is not “German vs English”, it’s ambiguity and noise:

• unstructured histories (e.g., “pt unwell since days, maybe fever?, meds??, unclear past medical history”)

• inconsistent terminology (e.g., mixing “Luftnot”, “dyspnea”, “can’t breathe” without onset or severity; “renal failure” without stage or creatinine)

• copy-pasted patient messages (e.g., long chat text: “I feel weird… heart skipping… since yesterday… btw could I be pregnant??”)

• Mixed languages in one prompt (e.g., “Anamnese: seit 3 Tagen Fieber. Triage: douleur thoracique 7/10. DDx: PE vs pneumonia. Angehörige: ‘sie isch hüt mega komisch gsi’.”)

The Swiss trap: “copy-paste soup” (note fragments + patient messages + dialect) is not a prompting strategy.
Rule: normalize, structure, and use clinical terms before asking for reasoning.

4) Swiss-specific tools are emerging (with caveats)

• AlpineAI’s “SwissGPT”: AlpineAI states SwissGPT can be used in all Swiss national languages and in English. But still, “Language support” is not the same as “clinically validated performance.”

• Apertus (EPFL/ETH/CSCS, Swiss AI Initiative): Apertus is presented as a large-scale open, multilingual Swiss model trained on 15 trillion tokens across 1,000+ languages, including Swiss German and Romansh. At the same time, Apertus is primarily a general-purpose LLM, and its medical usefulness depends on domain adaptation, clinical evaluation, and appropriate governance.

Bottom line

• Routine work: local language is fine.

• Complex reasoning: English is often reasonable, not guaranteed better.

• Always required: clinical judgment, verification, and missing-data checks.

LLM outputs are decision support, not decisions.

Subscribe now

Share

Sources

• Hamad AA. Medical research production in native languages: descriptive analysis of PubMed. Qatar Med J. 2024. PubMed: https://pubmed.ncbi.nlm.nih.gov/38746849/

• Alonso I, Oronoz M, Agerri R. MedExpQA: Multilingual benchmarking of LLMs for medical QA. 2024. arXiv: https://arxiv.org/abs/2404.05590 ; journal: https://www.sciencedirect.com/science/article/pii/S0933365724001805

• AlpineAI. SwissGPT language support (Safety page). Accessed 2026-01-19. https://alpineai.swiss/en/safety/

• Swiss AI Initiative. Apertus (official project page). Accessed 2026-01-19. https://www.swiss-ai.org/apertus

• ETH Zurich. Apertus press release. 2025-09-02. https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html

Below is a curated overview of Swiss-developed or Swiss-governed clinical AI tools where vendors explicitly state that data is processed and hosted in Switzerland and aligned with Swiss data protection law (nDSG/FADP). The focus is practical deployability in daily clinical work, not marketing claims.

But what does “Swiss-compliant” mean?

In this article, Swiss-compliant means:

• Swiss company or Swiss public institution behind the tool

• Data processed and hosted in Switzerland

• Explicit statement that user inputs are not used for model training

• Data Processing Agreements (DPAs) aligned with nDSG (and GDPR where applicable)

This significantly lowers legal and institutional risk. It does not replace individual verification.

🧠 Chatbots for clinical assistance and triage

SwissGPT / Swiss HealthAssist

by AlpineAI (https://alpineai.swiss)

in partnership with HIN (https://www.hin.ch)

Often described as the Swiss medical ChatGPT, SwissGPT supports drafting notes, summarizing patient histories, structuring reports, and assisting with clinical reasoning within a Swiss healthcare security framework.

Key points

• Swiss-hosted processing in ISO-27001-certified data centers

• Encrypted, transient data handling

• Explicit no-training-on-user-data policy

• Integration into HIN workflows for ambulatory care

Cons

• Pricey compared to free or low-cost alternatives, which may be sufficient for simpler use cases, starting from 89 CHF/month for a standard subscription.

Dr. Nuts

by In a Nutshell (https://inanutshell.ch)

Dr. Nuts is a guideline-based clinical decision-support tool for Swiss primary care. It focuses on curated medical knowledge rather than free-text documentation or transcription.

Key points

• Swiss company, built for Swiss GPs

• Guideline-driven clinical support

• Lower data-protection risk profile as it does not primarily process patient data

• Free to use for physicians (Login with GLN-Number required)

Cons

• Scope intentionally limited to decision support and knowledge retrieval

ConfIAnce

by Hôpitaux universitaires de Genève (https://www.hug.ch)

developed with Liip (https://www.liip.ch)

ConfIAnce is a patient-facing triage and information chatbot developed within a Swiss university hospital setting.

Key points

• Swiss public-sector governance and hospital data-protection standards

• Patient-facing. No login required, no identifiable patient data stored. Queries anonymized and processed on Swiss servers

• Free to use for the public and institutions

Cons

• Interface in French only, but accepts input possible in multiple languages.

✍️ AI scribes and documentation tools

MPAssist

MPAssist focuses on transcription and structured documentation for consultations, ward rounds, and therapy sessions. It is used across Swiss rehabilitation centers, specialty clinics, and long-term care facilities, including multiple Uroviva sites.

Key points

• Established use in Swiss healthcare institutions

• Swiss-hosted infrastructure

• DPAs aligned with Swiss data-protection requirements

• Strong focus on documentation efficiency

Cons

• Effectiveness depends heavily on workflow integration

• Requires careful setup, consent processes, and institutional alignment

Evoya AI

by Evoya (https://evoya.ai)

Evoya AI converts voice-based consultations into structured medical notes, with a focus on primary care workflows.

Key points

• Data processed exclusively in Swiss ISO-27001-certified data centers

• Explicit no-external-training policy

• Designed for general practitioner workflows

Cons

• Limited publicly available information on scale of deployment

Subscribe now

Share

Important disclaimer

Even when tools are Swiss-hosted and marketed as compliant, physicians remain responsible for verifying suitability in their own setting. Before using any AI tool, always check:

• the Data Processing Agreement and listed subprocessors

• data retention and deletion policies

• canton-specific KV or institutional requirements

• whether identifiable patient data is processed

Swiss infrastructure reduces risk. It does not remove accountability.

LLMs (Large Language Models)

• Generate text by calculating statistically likely word sequences, not by understanding meaning.

• Models such as ChatGPT, Claude, or Gemini are general-purpose language models, not medical systems.

ML (Machine Learning)

• Learns patterns from existing data and therefore inherits errors and biases present in those data.

• Performs only as well as the data on which it was trained.

NLP (Natural Language Processing)

• Enables software to recognise, analyse, and structure unstructured medical text.

• Forms the basis for allowing computers to “read” clinical notes and extract relevant information, for example for coding, quality analysis, or searching large text corpora.

ACI (Ambient Clinical Intelligence)

• Captures conversations between clinicians and patients and automatically generates draft documentation.

• The generated text often appears highly plausible but may contain subtle inaccuracies.

• Commonly used to pre-draft anamnesis and progress notes.

CDSS (Clinical Decision Support Systems)

• Provide alerts or warnings about potential risks without making decisions themselves.

• Typical use cases include drug–drug interaction warnings, risk alerts, or reminder systems.

Hallucination

• Refers to the generation of factually incorrect content that is linguistically convincing and therefore difficult to detect.

• Example:

  “According to the ESC Guideline 2023, ivabradine is recommended as first-line therapy in stable angina.”

  → Sounds credible, but is fabricated.

Algorithmic Bias

• Occurs when training data underrepresent or distort certain patient groups.

• Leads to systematically poorer model performance for specific populations.

Model Drift / Dataset Shift

• Describes declining model performance as data or clinical contexts change over time.

• Causes include new guidelines, shifting patient populations, or changes in clinical workflows.

Human-in-the-Loop

• Means that AI outputs are reviewed by a human before being used.

• Considered a minimum standard for clinically relevant AI applications.

Data Protection and Regulation (Switzerland)

• revDSG (Revised Federal Act on Data Protection): Governs how patient data may be collected, processed, stored, and shared.

• DSFA (Data Protection Impact Assessment): Required when introducing systems that pose increased risks to patients’ rights, particularly AI tools processing sensitive health data.

• GDPR (General Data Protection Regulation): Relevant when EU-based providers, servers, or cross-border data processing are involved.

  Subscribe now

The greatest time burden in daily clinical practice does not come from medical decision-making, but from documentation, phone calls, and administrative work. This is exactly where current AI-based assistant tools are being used: quietly in the background, clearly limited in scope, and without interfering with clinical responsibility.

This brief overview focuses on a small number of practical tools that address common bottlenecks and can be meaningfully integrated into everyday medical workflows.

Documentation & Clinical Text Work

Heidi Health

Uses Ambient Artificial Intelligence (Ambient AI) to capture clinician–patient conversations in the background and generate structured medical documentation, often in SOAP (Subjective, Objective, Assessment, Plan) format.

Useful for: reducing after-hours documentation, creating structured visit notes, maintaining consistent clinical text.

Arztbrief.ai

A focused tool for generating structured medical letters from keywords or speech input.

Useful for: speeding up standardized referral or discharge letters and reducing repetitive writing.

Dragon Medical One (Nuance)

Medical speech recognition software with AI-trained, specialty-specific language models.

The AI component supports context-aware medical dictation, not ambient listening.

Useful for: fast and accurate dictation for clinicians who prefer speaking over typing.

Front Desk & Phone Support

Fonio

An AI-based phone assistant that automatically answers incoming calls and structures requests for the practice team.

Note: In Switzerland, AI-based phone assistants are still at an early stage. Individual solutions such as Fonio are already being used, while many other providers currently focus primarily on the German market.

Useful for: reducing call volume at the front desk, minimizing interruptions during consultations, and improving patient accessibility.

Summary

AI assistant tools deliver the most value when they simplify administrative routines and free up time for clinical work. Today, their most practical impact is seen in documentation and practice organization rather than in medical decision-making.

Outlook: SwissMed AI

At SwissMed AI, we will be testing selected tools individually and sharing practical, real-world assessments. Further articles, tool reviews, and experience-based insights will follow.

Note: Before implementing AI tools, practices should conduct their own data protection impact assessment (DPIA) and review data processing agreements. This article does not constitute legal advice.

AI won’t replace residents.
But residents who use AI thoughtfully will work very differently from those who don’t.

As a resident today, this is how I’d realistically use AI on a normal workday — and where I’d draw a hard line.

I’d use AI for these 5 things

1. Drafting discharge summaries and clinic letters
This is also helpful for foreign-trained doctors.
AI can improve structure, clarity, and medical German — but I’d still review and edit everything before signing.

2. Looking up clinical information
As discussed in a previous issue, tools like AMBOSS AI and OpenEvidence are already part of daily practice.
They’re useful for focused questions, guideline reminders, and evidence summaries — faster access, same responsibility.

3. Translating medical language into patient language
This matters for national as well as foreign patients. AI helps adapt explanations linguistically and culturally, improving understanding and adherence.

4. Preparing handovers and case presentations
Turning complex cases into a clear narrative is a core clinical skill — AI helps refine, not replace it.

5. Checking my clinical reasoning for blind spots
Similar to what I described with GLASS AI in a previous post:
not for answers, but to challenge thinking — red flags, alternatives, “what am I missing?”

I’d avoid AI for these 3 things

1. Diagnosing patients
Pattern recognition without clinical context is risky.
AI doesn’t examine the patient — you do.

2. Drug dosing or treatment decisions
Small errors here have big consequences.

3. Anything involving identifiable patient data
If you hesitate even briefly, don’t paste it.

My rule on the wards: keep AI use C.L.E.A.R.

Before using AI, I’d run through one quick mental check:

C — Context
Am I using AI for information, language, or structure — not decisions?

L — Liability
Could I stand by this output in front of a senior, a patient, or in court?

E — Exclude
Have I excluded all patient identifiers?

A — Authority
Am I treating AI as support, not as the authority?

R — Review
Have I read and verified every line before using it clinically?

If it isn’t C.L.E.A.R., I wouldn’t use it.

Final thought

Good AI use doesn’t save time by cutting corners.
It saves time by protecting clinical judgment.

Used well, AI doesn’t replace thinking —
it creates space for it.

1. OpenEvidence – Evidence in Seconds
   Scans guidelines and systematic reviews to give concise, referenced summaries in seconds.
   Use it for: quick evidence checks during ward rounds or consults, and rare/complex conditions.​

2. AMBOSS.com AI

   AMBOSS is tailored to DACH clinical practice and exams and is expanding its AI features.

   Use it for: fast information lookup, differentials, algorithms, and clinical pearls.

   At many universities and teaching hospitals, AMBOSS access is included via institutional/campus licenses, so students and clinicians can use it at no additional personal cost.

   AMBOSS has also announced an “AI Mode (LiSA 1.0)” for the German Version focused on medical- and nursing-specific information research (Status: 28 Jan, 2026).

3. UpToDate AI – Guideline-Backed Answers
   Natural-language queries on top of UpToDate’s content, returning focused, linked summaries.
   Use it for: confirming management plans, therapy choices, and guideline-compliant decisions.​

4. Glass AI – Reasoning and Differentials
   Helps structure assessment and plans and expand differentials from your clinical description.
   Use it for: complex cases, atypical presentations, and improving case presentations and notes.​

Why this matters

Used selectively for tricky cases, admissions, and discharges, these tools compress hours of searching into minutes, freeing time for direct patient care, supervision, and teaching—while keeping your clinical judgment firmly in charge!

Quick comparison

Switzerland has introduced clear policies to ensure artificial intelligence systems operate transparently and responsibly, balancing progress with strong privacy safeguards. These measures require organizations to clearly inform individuals when AI influences decisions about them and explain how the technology functions.

Key points include:

• Individuals have the right to challenge AI decisions and request human involvement to ensure fairness.

• High-risk AI systems, particularly in sectors like healthcare and finance, must be supported by comprehensive documentation, risk evaluation, and continuous oversight.

• Existing data protection laws are fully applicable to AI, guaranteeing secure and purpose-limited data use.

• Switzerland plans to adopt the Council of Europe’s AI Convention, promoting ethical AI use and increased transparency.

• Rather than forming a new AI regulatory authority, existing sector-specific bodies are tasked with ensuring AI compliance.

• Upcoming regulations will include clearer labeling of AI-generated content and restrictions on unethical AI surveillance.

This approach aims to cultivate trust in AI technologies by making them understandable, accountable, and safe, while also preserving Switzerland’s leadership role in AI innovation.

Sources:
Federal Data Protection and Information Commissioner, 2025
Bonnard Lawson, Swiss AI Regulation Overview, 2025
Swiss Federal Council AI Strategy, 2025
Lenz & Staehelin Legal Insights, 2025