|
AI is reshaping healthcare fast. Below are 3 key AI developments, 2 studies, and 1 takeaway for this week to help you better lead with AI. Target read time: 5 minutes.
|
On June 30, Anthropic released Claude Science in beta, a workbench that runs all stages of research in one environment: literature analysis, data processing, figure generation, and manuscript prep. It ships with 60+ preconfigured skills spanning genomics, proteomics, and cheminformatics, renders 3D protein structures and genome tracks natively, and includes a reviewer agent that checks citations and calculations. Every analysis produces an auditable artifact with its complete code history. Early users include the UCSF Brain Tumor Center.
My Take
Reproducibility is a key part of serious research. Therefore, components like the reviewer agent and the auditable code history are built specifically for work that has to survive peer review. At the same time, this again feels like Claude keeps eating into the app layer.
Read the announcement →
On July 1, Iowa's HF 2635 let utilization‑review organizations use AI for initial prior‑authorization reviews while barring it from being the sole basis to deny, delay, or downgrade a request. Indiana and Tennessee laws took effect the same day: insurers can't downcode claims on AI alone, and AI systems can't advertise themselves as licensed mental‑health professionals. Washington is allowing AI to approve prior auths while reserving denials for licensed clinicians. Georgia and Utah follow on January 1, 2027.
My Take
Previously, this state‑by‑state AI decision‑making looked like a patchwork. While that hasn't changed, there's a common principle emerging: AI can say yes, only a human can say no.
Read the analysis → | Read the state tracker →
On June 24, Prosus announced a €480 million round in Alan, the French health insurer, at a €5.5 billion valuation. Alan covers 1.1 million members across France, Spain, Belgium, and Canada, passed €800 million in annual recurring revenue in Q1 (up 53% year over year), and is profitable in its home market. It runs insurance, prevention, and care delivery on one platform, and the round funds AI‑led product development and international expansion.
My Take
Most US payers are retrofitting AI onto incumbent, decades‑old systems. In contrast, internationally, this is a test worth monitoring of what an insurer designed around AI can do when prevention, care delivery, and the insurance risk all sit in the same P&L.
Read the announcement →
|
A pragmatic cluster‑randomized trial embedded a GPT‑4o‑based "AI Consult" tool in the medical record at 16 primary‑care clinics in Kenya, covering 9,691 patients seen by 103 clinicians. The tool flagged likely diagnostic and treatment gaps as the visit happened. Treatment failure at 14 days was statistically no different: 2.2% with the AI versus 2.0% without. Documentation, diagnoses, and treatment plans all measurably improved, and the tool was safe.
My Take
This is an interesting trial at this scale to test whether AI changes what happens to patients (not just the impact on clinicians). Lots of important limitations, yet — this is an important null result.
Read the study →
In a cluster‑randomized trial at 2 California emergency departments, 66 attending physicians received near‑real‑time, LLM‑generated feedback on their compliance with the CMS sepsis bundle (SEP‑1), replacing the usual delayed, sampled chart review. Across 301 patient encounters, compliance rose from 70.1% to 82.9%, a 13‑point gain, driven mostly by better documentation of fluid administration. Yet, 30‑day mortality and ICU admissions didn't change.
My Take
Hospitals spend real abstractor hours measuring SEP‑1 today, so the automation likely pays for itself before any behavior change. The 13‑point gain came from making feedback immediate. That rapid loop is ideally reusable across nearly every quality measure.
Read the study →
|
The process improved. The patient didn't notice.
Two randomized trials of clinical AI published this week both worked, just not where it counts most. In Kenya, GPT‑4o decision support made documentation and diagnoses measurably better across 9,691 patients, but treatment failure didn't budge. In San Diego, AI‑generated feedback lifted sepsis bundle compliance 13 points, yet mortality didn't move either.
These aren't weak studies; they're important and relevant. The gap is in what we measure and optimize for. When a process metric carries the outcome inside it, moving it moves patient metrics. For example, the classic ICU checklist study cut catheter infections by up to 66% because every checklist item was itself a causal step in driving safety.
Takeaway
The tools are clearly ready to move process measures. The ideal next wave of trials aims them at endpoints that carry the outcome inside them, like time to antibiotics or missed diagnoses found, where moving the metric is the same thing as helping the patient.
|
|