Welcome back to Dose of AI — the newsletter that cuts through the noise to bring you the healthcare AI stories that actually matter. This week: the WHO drops a landmark report on AI across Europe, Nature Medicine asks an uncomfortable question about whether AI is actually helping patients, and doctors are starting to code their own clinical tools. Let's get into it.

Published: April 20, 2026 | Source: WHO/Europe

The World Health Organization's European office released what it's calling the first comprehensive review of AI adoption across all 27 EU Member States — and the headline is that Europe is moving fast, but not recklessly. The report, built on data collected between June 2024 and March 2025, found that nearly three-quarters of EU countries are already using AI-assisted diagnostics, including tools that support medical imaging, disease detection, and clinical decision-making. Meanwhile, 63% have deployed patient-facing chatbots, and nearly half of member states have created dedicated professional roles for AI and data science in health systems.

The report also flags where work remains. Workforce training hasn't kept pace with deployment — and while 81% of EU countries involve stakeholders in AI governance, patient and public consultation is still inconsistent.

Why It Matters: This is more than a progress report — it's a policy baseline. For the first time, there's a shared, comparable picture of where 27 different health systems stand on AI. That matters enormously for health equity. If AI tools are trained and evaluated on narrow population data, their benefits won't reach everyone equally. The WHO is clearly pushing to make governance and literacy non-negotiables, not afterthoughts. For U.S. health systems watching this space, Europe's framework-first approach offers a useful contrast to the American innovation-first model.

Published: April 21, 2026 | Source: Nature Medicine

In a pointed policy piece published this week, researchers Anna Goldenberg and Jenna Wiens posed a challenge to the healthcare AI industry: all this investment, all these deployments, all these bold claims — but is AI actually improving patient outcomes? Their answer is unsettling: in many cases, we simply don't know.

The authors identify two core problems holding back honest evaluation. First, the industry leans heavily on accuracy metrics and model performance scores that don't necessarily translate into meaningful clinical improvement. Second, when outcome metrics are used, they often capture factors unrelated to what the AI actually contributed. Predictive models flag at-risk patients, ambient scribes draft notes, computer vision triages scans — but attribution remains murky. The hype, they argue, has outpaced the evidence.

Why It Matters: This piece is a necessary corrective. The healthcare AI market is projected to reach extraordinary scale over the next decade, and hospitals and health systems are making real purchasing decisions based on vendor promises of efficiency gains and burnout relief. Without rigorous, outcome-focused evaluation frameworks — not just accuracy benchmarks — it's hard to separate tools that genuinely help patients from those that simply generate impressive demos. Expect this paper to ripple through procurement conversations and regulatory discussions in the months ahead.

Published: April 24, 2026 | Source: Healthcare IT News

A fascinating development broke today: physicians are increasingly turning to agentic AI coding tools to build custom clinical applications — without traditional software development backgrounds. At a webinar hosted by Anthropic Thursday, emergency medicine physician Dr. Graham Walker and interventional cardiologist Dr. Michał Nedoszytko demonstrated how they're using tools like Claude Code to create patient care and workflow applications from scratch. "If the EHR is a problem, maybe just create your own," Nedoszytko said — a line that's already circulating widely in health IT circles.

The potential is real: clinician-built tools could close the long-standing gap between what technology can theoretically do and what clinicians actually need. But security experts are sounding the alarm. With AI models now capable of detecting software vulnerabilities, healthcare organizations face a new class of cyberattack risk — and clinician-built apps may not undergo the same security audits as enterprise software. The Cloud Security Alliance released a whitepaper this month urging every health organization to build a 90-day AI cybersecurity preparedness plan now.

Why It Matters: The idea of doctor-built software might sound like a niche curiosity, but it signals something much larger: a structural shift in who controls clinical technology. For decades, health systems have been at the mercy of large EHR vendors and slow IT procurement cycles. If clinicians can build the tools they actually want to use, care delivery could improve faster — and more relevantly — than any vendor roadmap allows. The catch is governance. A tool built by a brilliant cardiologist over a weekend is not the same as an FDA-cleared device. Health systems need frameworks for this now, not after the first incident.

Dose of AI is an independent weekly briefing on healthcare artificial intelligence. All views are analytical and do not constitute medical, legal, or investment advice.

Sources: