The week in medical AI, judged by one question: does this actually change what we do for patients? Saturday 27 June 2026.

• Diagnostic AI: EchoNext, among the first multi-condition cardiology AIs cleared by the FDA (23 June), reads a standard 12-lead ECG to flag six forms of structural heart disease. In its Nature validation it caught 77% of structural heart problems versus 64% for 13 cardiologists reading the same 3,200 ECGs, but it triages, it does not diagnose, and the echo still decides.
• Ambient AI, a reality check: in 263 clinicians across six health systems (JAMA Network Open), burnout fell from 51.9% to 38.8% in 30 days, yet time savings are modest on average and concentrate in heavy users, while the value shifts into the revenue cycle.
• The skeptic's corner: a new Nature paper shows medical-AI models can leak whether an individual patient was in the training data almost perfectly, even when aggregate privacy metrics look safe, and underrepresented patients are the easiest to single out.
• From the literature, via PubMed: two big meta-analyses on AI reading scans (breast nodal status, AUC 0.84; rectal complete response, AUC 0.85) with likelihood ratios that say adjunct, not replacement.
• Follow the money: once Medicare paid for AI to detect large-vessel strokes (NTAP), billed use reached about 21% by 2022, but adoption tracked the hospital, not the patient, an equity problem (AJNR).
• Teaching point: compared with what? Four questions before any tool changes a decision.

A new episode every Saturday. Subscribe, rate the show, and send your feedback.

Sources & further reading: Nature; Nature (privacy audit); European Radiology; Molecular Imaging; American Journal of Neuroradiology; JAMA Network Open; FDA; NewYork-Presbyterian / Columbia; PubMed and the National Library of Medicine.

The week's most important developments in medical AI, and the one question that matters: does this actually change what we do for patients? Saturday, 20 June 2026.

In this episode:

• Diagnostic AI. Hetairos predicts 102 methylation-defined CNS tumour subtypes from a routine H&E slide. Built and validated on 9,606 patients and 11,000+ slides across 11 centres on 4 continents; from histology alone it scored 0.87 on confident calls, and 0.68 versus 0.30 for five board-certified neuropathologists, turning a roughly 12-day molecular workup into about 12 minutes. Why molecular testing still rules, and where this really changes access.

• Ambient AI grows up. Abridge expands beyond the scribe into coding, prior authorisation, claims and decision support, backed by a strategic Eli Lilly investment. Philips' Future Health Index 2026: 46% of clinicians save at least 132 hours a year, half report capacity for about 8 more patients a week, and 65% increased their AI use at work.

• The skeptic's corner. Overtrust in AI medical advice, how the way you prompt can steer a model toward more accurate but also more harmful answers, and the habit that protects you: compared with what?

• From the literature, via PubMed. The AMIE RCT in Nature Medicine (assisted care preferred 47% vs 33%, fewer clinically significant errors 13% vs 24%), a 70-clinician RCT in npj Digital Medicine, plus the PROTEUS AI stress-echo trial.

• Follow the money. CPT 2026 adds 288 new codes including AI services (live since 1 January), NHS England commits £20m to scale AI chest X-ray, the US HHS issues an RFI on AI to cut costs, and the WHO publishes a discussion paper on AI in health policy.

• The teaching point. A performance metric is a signal, not proof of benefit. A number is a signal, not a target.

A new episode every Saturday. Subscribe, rate the show, and send your feedback.

Sources & further reading: Nature Cancer (Hetairos); Digital Health News and Philips Future Health Index 2026; NEJM AI and Lancet Digital Health; Nature Medicine and npj Digital Medicine (via PubMed / National Library of Medicine); AMA CPT 2026; NHS England; US HHS; WHO.

This week’s AI in Healthcare update covers four key developments: an NEJM AI randomized trial finding AI literacy training did not prevent automation bias from intentionally erroneous LLM diagnostic suggestions; FDA clearance and a CMS reimbursement pathway for Bunkerhill’s AI tools quantifying coronary and aortic valve calcium on routine contrast, non-gated chest CT (with limited independent peer-reviewed validation noted); a 48-trial Bayesian network meta-analysis (34,106 participants) reporting five AI colonoscopy systems improved adenoma detection rate but not advanced adenomas or sessile serrated lesions; and a Nature Medicine perspective urging higher evidentiary standards and patient-centered outcomes before claiming AI improves healthcare.

00:00 Weekly AI Healthcare Briefing

00:48 AI Literacy vs Automation Bias

02:22 FDA Cleared Calcium Detection

03:55 AI Colonoscopy Meta Analysis

05:54 Raising the Evidence Bar

07:23 Wrap Up and Next Week