About

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.