THE MED-AI BRIEF
{{current_date_full}} · 5-MINUTE READ · Read online
Good morning, {{first_name|Doctor}}. A US insurer says hospitals billed 55,158 more inpatient stays as "complex" than they did in 2023, without treating those patients any more intensively, and names ambient listening among the tools finding the extra diagnoses. The NHS is rolling out that same class of tool right now.
In today's Med-AI Brief:
🩺 Research: the scribe that listens also writes the bill
⚖️ Regulation: Ireland's health service sets five gates for AI
🌍 Use cases: this week in Riyadh, the SFDA on AI devices
🛠️ Practice: read the diagnoses you did not write
RESEARCH
🩺 The scribe that listens also writes the bill
The brief: On 24 September the Blue Cross Blue Shield Association published a white paper on hospital coding intensity, estimating $942m in extra costs to its member plans over two years. The paper argues this is not more operations or sicker patients. It is more secondary diagnoses on each claim, pushing stays into higher-paying categories, and it says the rise "coincides with hospitals adopting new AI-assisted billing software that scans records and lab data for anything that can be coded."
The details:
The share of inpatient cases billed as complex rose from roughly 37% in early 2023 to about 40% by late 2025. Against the 2023 baseline that is 55,158 extra complex cases, worth $653m at an average of $11,800 each.
The fastest-growing "bump codes" are unspecified acidosis, hyponatraemia and acute posthaemorrhagic anaemia. In the paper's words, "many of these diagnoses can be derived from single laboratory values," and it names "ambient listening for observation codes and laboratory data mining" as likely contributors.
The discordance test: hospitals with the steepest coding growth diagnose anaemia 38% more often than their peers, yet transfuse fewer of those patients (16.9% against 19.3%).
Why it matters: Read the claim and the evidence separately. This is an insurer analysing its own claims, not a peer-reviewed study, and it does not show which hospitals actually used AI tools; its own wording is "coincides" and "may be linked." The American Hospital Association disputes it, pointing to an ageing population and milder cases moving to outpatients. But the mechanism does not need the US. In England, 2026/27 NHS Payment Scheme unit prices are "used for the variable element of aligned payment and incentive arrangements, and for activity-based payments," and HRG4+ scores complications and comorbidities from secondary diagnoses. NHS Digital's curated CC lists blunt the effect, since not every extra code moves the HRG. They do not change who signed the note that the coder read.
REGULATION
⚖️ Ireland's health service sets five gates for AI
The brief: On 23 September Ireland's Health Service Executive published an AI Implementation Framework: a five-phase route every AI project must follow, from first idea to monitoring after go-live. It is designed to deliver the Irish government's AI for Care strategy, published earlier this year.
The details:
The five phases: opportunity identification and prioritisation; planning and approval; design and readiness; integration and deployment; operations and monitoring.
Phase one requires "a preliminary assessment for clinical safety, ethical considerations, regulatory compliance," scored on the HSE risk matrix. A clinical advisory group will "manage clinical safety issues and integrate AI into clinical workflows."
Patient and public representatives are to be "actively involved" in governance groups, and clinical projects carry "benefits realisation and post-market surveillance" after deployment.
Why it matters: This is a process document, not law, and the published summary says little about who carries liability when a tool gets it wrong. But it answers the question a ward actually needs answered: who checks the tool before it arrives, and who keeps checking afterwards. In England the equivalent answer is still pending, with the government response to the National Commission's 44 recommendations unpublished three issues on. Any trust or Gulf hospital drafting its own AI policy could borrow the five gates this week.
USE CASES
🌍 This week in Riyadh: the SFDA on AI devices
The brief: Saudi Medtech 2026 runs 29 to 30 September at the Radisson Blu Hotel Riyadh Convention and Exhibition Center. It is invitation-only and registration has closed, but the agenda puts the regulator, the ministry and hospital CIOs on the same stage, and that is worth reading on its own.
The details:
Day one: Dr Bandar Almifgai, head of digital health at the SFDA, on "SFDA's Regulatory Approach to AI Medical Devices and Digital Health Innovations", and a panel on "AI Governance & Patient Safety: Creating a National Framework for Responsible AI", moderated by SDAIA AI adviser Dr Metab AlKubeyyer.
Also on day one: "Scaling AI in Hospitals: From Pilots to Predictive and Autonomous Systems", with CIOs from Fakeeh Care and KSU Medical City.
Day two is imaging: RSSA president Dr Arwa Badeeb on "Radiology 2030", and Dr Amna Kashgari of the Ministry of National Guard Health Affairs on "Beyond Algorithm: AI Governance in Radiology".
Why it matters: An agenda is not an announcement, so treat this as a watch list. The SFDA session matters because the SFDA decides which AI devices reach Saudi wards. The governance panel matters because its title asks the question the UK Commission spent a year on: who is accountable when the tool is wrong. If either says anything new on the record, it will be in next week's brief.
PRACTICE
🛠️ Two minutes: read the diagnoses you did not write
Next time you sign an AI-drafted note, skip the prose for a moment and read only the diagnoses and problem list. For each one, ask a single question: did I assess this, and would I act on it?
A mildly low sodium you glanced at is not the same as hyponatraemia you are managing. If the scribe or a coding prompt has turned an observation into a diagnosis, edit it before you sign. It is quicker than explaining it to an auditor, and more honest to the next clinician who reads the record.
QUICK HITS
📰 Everything else
Scribes at scale: The US Department of Veterans Affairs picked Abridge for its ambient AI enterprise contract on 22 September, a multi-vendor deal with a ceiling of $775.7m over five years. The VA reports more than 986,000 primary care appointments already using ambient scribes.
Regulatory science: The MHRA, Medical Research Council and Office for Life Sciences launched £20m for four regulatory science centres on 22 September, one dedicated to data-driven approaches and AI. Applicant webinar on 13 October.
Scotland: The Programme for Government 2026-31, published 22 September, commits to an AI hub "to accelerate the safe adoption of AI technologies across the system." No detail yet on who runs it or what it assures.
Still counting: DHSC and the MHRA have not published their formal response to the National Commission's 44 recommendations. Third issue running.
Watching: HSSIB's investigation into ambient voice technology in hospitals. No reporting date given.
How was today's brief?
✉️ One favour: hit reply and tell me whether anyone at your trust or hospital has told you what happens to your AI-drafted notes after you sign them.
Until next week,
Saeed
Know a colleague who signs AI-drafted notes? Send them your referral link.
Nothing here is clinical, legal or coding advice. Check any tool against your own information-governance rules before it touches patient data.