Issue 9 · 8 September 2026 · Fortnightly | AI in Healthcare · Australia & Beyond
This fortnight starts close to home. A University of Queensland study finds that nearly 80% of mental health clinicians are already using AI to listen in on and transcribe sessions — well ahead of any formal guidance on how. In the US, a bipartisan bill would stop AI from being the thing that decides a claim isn't medically necessary, which is worth reading against every NDIS, DVA and Medicare knockback you've ever appealed. On the diagnostics side, breast screening AI has crossed a genuine line — reporting some mammograms as normal with no radiologist looking at all — and a 30-second video of someone's face can now flag diabetes and high blood pressure without a cuff or a blood draw.
Here's what you need to know:
Summary
Read time: ~4 min
Top Stories
- A University of Queensland study finds AI use has outpaced governance in mental health practice
- A bipartisan US bill would bar AI from making healthcare coverage determinations without a licensed professional
- Vara's autonomous AI triage for breast screening earns a world-first CE mark
- AI reading facial and palm video detects hypertension and diabetes without a cuff or a blood draw
Signals
- NDIS social and community participation budgets are cut by roughly half from next month
- AI paired with a microfluidic chip cuts bacterial identification from days to hours
- The next round of global AI-health rules is being written this month in Hangzhou
- A new framework sorts clinical AI into three tiers of autonomy — useful for reading a vendor pitch
Top Stories
Nearly 80% of Mental Health Clinicians Are Already Using AI. The Guidelines Aren't.
🇦🇺 Australia | Practice Standards
The study, reported 3 September, was led by PhD candidate Benjamin Johnson at UQ's National Centre for Youth Substance Use Research. It set out to establish how mental health clinicians actually use AI day to day, not how peak bodies assume they use it — and the gap between the two is the finding.
AI performed well for the administrative load clinicians handed it: transcription, summarising, drafting routine correspondence. It performed substantially worse the moment a case stopped being routine. Johnson's framing was direct — AI is a suitable supplement for administrative work, but "it's not in a position to replace clinicians when dealing with complex cases that require accurate and safe diagnosis and treatment recommendations."
No professional body has yet published binding guidance that matches how widely the tools are already being used. That leaves individual clinicians deciding, case by case and largely alone, where the administrative-help line ends and the clinical-judgement line begins — precisely the ambiguity a formal standard exists to remove.
What it means for clinicians:
- Write down, for your own practice, exactly where you stop using AI output as-is and start checking it against your own judgement. If you can't state the line clearly, you don't have one yet — you have a habit.
- Treat transcription and drafting as the safe zone and anything touching risk assessment, differential thinking or treatment recommendations as the zone that gets read twice, by you, before it goes anywhere near a file.
- Ask your peak body directly whether formal AI guidance is coming and when. A clear answer protects you if a notification ever turns on how you used the tool, not just on what it produced.
Congress Just Tried to Stop AI From Deciding What's "Medically Necessary"
🌏 Global | Practice Admin
The bill, introduced by Reps Greg Landsman, Buddy Carter, Kim Schrier and Tom Barrett, targets the specific point where an algorithm currently gets to say a claimable service doesn't qualify. It would require disclosure whenever AI is used in a coverage notice, and it would treat AI-driven utilisation review of mental health claims as a treatment limitation under parity law — meaning it can't be used to apply a tighter standard to mental health than to anything else.
This is a US insurance fight, not an Australian one, and no equivalent bill exists here. But the underlying mechanism — software assessing whether a support is "reasonable and necessary" before a human ever looks at the file — is exactly the shape NDIS, DVA and Medicare assessment already takes, and the direction every funder is heading as automated triage gets cheaper to deploy.
The bill's premise is worth sitting with regardless of jurisdiction: a clinical determination about a person's care shouldn't be something a system generates and a human rubber-stamps. That is the standard to hold any funder to, not just the one this bill happens to regulate.
What it means for clinicians:
- Next time a submission comes back short, ask explicitly whether an automated tool flagged it before a human reviewed the file. You are entitled to know what actually assessed your client's claim.
- Write funding submissions assuming the first pass may be algorithmic, not human. Lead with the specific, measurable functional need — vague clinical language is what a pattern-matching system is worst at handling fairly.
- Watch whether NDIA or DVA disclose their own use of automated assessment tools. A funder that won't say how a decision was reached is a harder decision to appeal, whatever generated it.
The First AI Cleared to Read a Mammogram With No Radiologist Looking At All
🌏 Global | Diagnostics
The CE certification (EU Medical Device Regulation, Class IIb) covers autonomous triage within organised population screening. Exams the system classifies as clearly normal get reported by AI alone; anything else routes to a radiologist as usual. A real-time safety layer called ATMON tracks each site's performance, hardware and daily signals, and automatically reverts a site to full radiologist reading if anything drifts outside set limits.
The evidence behind it is substantial rather than promotional: a 2025 study in Nature Medicine followed 461,818 women with no exclusion criteria, the largest prospective study of AI in breast screening to date. That scale is what separates this from the accuracy claims Issue 8 warned you to be sceptical of — this is outcome-adjacent population data, not a lab benchmark.
BreastScreen Australia is already moving in the same direction. BreastScreen NSW introduced AI-assisted reading in November 2024, cutting radiologist workload by close to half, and a national trial involving around 200,000 women is now testing whether AI can safely support double reading here. Vara's certification is a preview of where that trial is heading, not a separate story.
What it means for clinicians:
- If you work with clients before or after breast cancer treatment — exercise physiology, lymphoedema management, post-surgical rehab — expect earlier and more consistent detection to shift referral timing over the next few years, not overnight.
- Use "autonomous" as a specific term, not a marketing word, when a vendor uses it on you. It should mean a defined subset of cases handled with no human review and a monitored safety mechanism — ask what happens when the system is uncertain, not just how accurate it is on average.
- If a client mentions a screening AI system to you, you can tell them accurately: it's reporting the clearly normal cases, and a radiologist still reviews everything else. That's a more precise answer than most people are getting from the coverage of this.
A 30-Second Video of Someone's Face Can Now Flag Diabetes and High Blood Pressure
🌏 Global | Diagnostics
Teams from the University of Tokyo and Institute of Science Tokyo recorded short, high-speed facial and palm videos on a spectroscopic camera and trained a machine-learning model on pulse-wave dynamics, skin blood-flow patterns and spectral skin colouring. From a 30-second recording, the algorithm detected hypertension with 95.0% accuracy; from just 5 seconds, accuracy was still 90.3%.
Diabetes detection worked the same way, reading facial blood-flow patterns instead of pulse waves: 88.2% accuracy from a 30-second video, 81.2% from 5 seconds. Lead researcher Ryoko Uchida described the goal plainly — contactless screening in everyday environments, at scale, for conditions that are common and frequently undiagnosed.
The appeal for allied health is the same one Issue 8 flagged for AI-read ECGs: no new hardware, no new workflow, just more information pulled from something already within reach. A telehealth consult with a camera is now, in principle, also a cardiometabolic screening opportunity.
What it means for clinicians:
- If you run telehealth consults for NDIS or DVA clients in regional or remote areas, this is the class of tool to watch — it could flag undiagnosed risk before you start an exercise program, without a referral round-trip first.
- It's a screening flag, not a diagnosis. Treat a positive result the way you'd treat any other red flag: a reason to refer for confirmation, not a reason to change a treatment plan yourself.
- This is a research presentation, not a cleared product on the ARTG. File it as something to ask about in twelve months, not something to expect a vendor to offer you next week.
Signals
🇦🇺 Australia | NDIS Funding — From 1 October, social, civic and community participation budgets reset to roughly 2023 levels — a cut of around 50% — phasing in over about 12 months as individual plans renew, with average spending in this category falling from around $31,000 to around $26,000 a year. Capacity-building daily activity budgets drop a further 10%. Nothing changes for a given client until their plan renews, so the date to diary is theirs, not 1 October. Read →
🌏 Global | Diagnostics — Published 21 August, a Southern Illinois University system pairs a microfluidic chip with a U-Net AI model to identify bacteria and their drug response from a blood sample in hours rather than the usual 24 to 48. Every hour of delayed effective treatment in the first six hours of sepsis raises mortality risk, so a faster read in hospital pathology is what eventually produces a more stable client, sooner, for the rehab clinicians who see them next. Read →
🌏 Global | AI Regulation — WHO, the ITU and WIPO convene the third meeting of the Global Initiative on AI for Health from 16 to 18 September, the forum that shapes the standards TGA and AHPRA guidance eventually draws on. Nothing binding comes out of this fortnight, but it's worth ten minutes with the agenda if you want to see a direction before your vendor's marketing catches up to it. Read →
🌏 Global | Vendor Claims — Published 1 September in Frontiers in Science, the framework describes a spectrum running from "advisory" tools that sit outside the workflow, through "copilot" tools that share the task with a clinician who stays in control, to "navigator" systems that run with minimal human oversight. Next time a vendor calls their product revolutionary, ask which tier it actually sits in — almost everything on the market today is still advisory. Read →