Senior correspondent, research & policy
Abdelmoumin reports on the evidence and the rulemaking behind medical AI: trial design, post-market surveillance, reimbursement, and the widening gap between what a model reports on a benchmark and what it does in a clinic. He reads the appendix so you do not have to.
26 stories · April 8, 2026 — August 4, 2026
A draft framework would let approved clinical models update on new data, in exchange for continuous performance reporting and a hard rollback obligation.
Job postings for the role have roughly tripled since the start of the year. What the position actually does varies enormously.
Eleven hospitals trained without moving data. The performance is respectable, the coordination cost was enormous, and the paper is unusually honest about both.
Turnaround times have collapsed. A rising share of denials cite reasoning that reviewers on both sides describe as difficult to argue with or against.
Researchers propose scoring clinical models on appropriate abstention, arguing that confident answers on out-of-scope cases are the dominant real-world failure.
Three acquisitions this quarter have moved narrow, well-validated tools inside platforms that price by the suite.
Ethicists argue that burying model disclosure in admission paperwork satisfies a legal requirement and defeats the purpose.
An audit found the strongest feature was distance to clinic, and the resulting overbooking policy fell hardest on the patients least able to absorb it.
A review of 340 cleared devices found continuous real-world performance monitoring in fewer than one in ten deployments.
When variant interpretation drops from weeks to hours, the bottleneck moves to the conversation about what to do with the result.
Generated notes score well on completeness rubrics designed for human writing. Chart review suggests the rubrics no longer discriminate.
A long-running mammography program reports its first durable change in recall rates — downward, after an initial rise that took eighteen months to reverse.
Three suits filed this spring turn on whether a clinician who followed a model's recommendation, or overrode it, met the standard of care.
A methodological review finds that fewer than one in five published evaluations measured a patient outcome rather than a model metric.
A security review found that standard de-identification left enough structure for a general purpose model to link records across datasets.
Generated summaries are complete and long. Receiving clinicians report missing the one sentence that used to matter.
Payment pathways exist for a handful of imaging tools and almost nothing else, and the gap is shaping what gets built.
Nearly two-thirds of surveyed systems now run subgroup performance audits. Fewer than a fifth have ever changed a deployed model because of one.
Legal teams are separating generated documentation from generated inference, and asking clinicians to keep the second one out of the chart.
Adenoma detection did not improve at high-performing sites. The subgroup where it did is the finding that matters.
Waitlists are long and the tools are cheap. The safety literature is limited to short trials in low-acuity populations.
Automation removed a documentation bottleneck that turned out not to be the bottleneck.
Models that read eligibility criteria work well. Models that read the patient's chart to check them do not.
Draft replies to patient portal messages were clinically sound and read, to patients, as though nobody had actually looked.
Six states have enacted requirements with materially different triggers. Compliance strategy has converged on writing one policy for the toughest.
Payers are adapting to provider-side coding models, and the resulting arms race is consuming attention on both sides.