🏥 RecovrSignal Issue 12: 8 of 12 AI scribes may be medical devices, and three AI wins from overseas

Issue 12 · 27 October 2026 · Fortnightly | AI in Healthcare · Australia & Beyond

The regulator has put a number on something practices have worried about for a year: 8 of 12 AI scribes it looked at have features that likely make them medical devices, and it will not say which. It was a quieter fortnight for local news, so the three wins that follow come from overseas: an AI that took patient histories before urgent care visits without a doctor needing to step in, a voice clip that may flag type 2 diabetes in 20 seconds, and an autonomous triage system that handled 94% of urgent skin cancer referrals at two UK hospitals.

Here's what you need to know:

Summary

Read time: ~4 min

Top Stories

• The TGA reviewed 12 digital scribes, found 8 with functions likely to make them medical devices, and flagged compliance action against suppliers
• A Lancet study of Google's AMIE found an AI could take a patient history before an urgent care visit with no conversation needing to be stopped
• Researchers from thymia and RMIT University told the EASD meeting in Milan that an AI can screen for type 2 diabetes from 20 seconds of speech
• A real-world study presented at EADV found AI triage let hospitals discharge benign skin lesions without review, with six missed cancers found later

Signals

• Mount Sinai researchers built a wearable AI that forecasts prolonged sitting in women with chronic pelvic pain
• The TGA will not name the 8 scribes flagged as possible medical devices, so practices have to ask their own vendors
• A colonoscopy study found doctors' detection rate fell from 28.4% to 22.4% when they went back to working without AI
• ARPA-H awarded the University of Vermont up to US$38 million to build AI digital twins for critically ill patients

Top Stories

The TGA Says 8 of 12 AI Scribes May Be Medical Devices, and It Won't Say Which

🇦🇺 Australia | AI Regulation

The TGA's starting point is that transcription alone does not make software a medical device, but features that suggest diagnoses or treatments can. Its review covered 12 digital scribe suppliers. None described their product as a medical device, yet the TGA found that 8 include functions likely to meet the definition in section 41BD of the Therapeutic Goods Act. Two products needed more information and two looked unlikely to qualify.

The review has limits that matter. The TGA did not test any products, check safety or performance, or assess privacy protections, so it is not saying these tools are unsafe. It will not publish product-specific findings either. The concerns it named were scope creep, AI that influences clinician decisions, and weak transparency and post-market feedback. After 12 months of engagement with vendors, the regulator says its focus is moving to compliance action against suppliers.

For a private practice this is an awkward place to be. One Newcastle GP has questioned the value of findings that flag a problem without saying which products are affected, and a practice owner told the same outlet that small practices cannot tell which scribes are safe to use. The practical reading is that the risk sits in features that go beyond writing up what was said, such as suggested diagnoses, treatment options or goal recommendations. If your scribe does that, find out who is responsible for it.

What it means for clinicians:

• Ask your scribe vendor in writing whether any feature suggests diagnoses, treatments or goals, and whether they have sought ARTG inclusion or been contacted by the TGA. Keep the reply with your practice records.
• If a suggestion feature is switched on and you do not need it, turn it off. Keep the scribe doing the job it was bought for: turning what was said into a draft note that you check and sign.
• Update your client consent wording so it names the scribe in use and says you review and sign every note. NDIS, DVA and Medicare clients are likely to start asking after this coverage.

An AI Interviewed Patients Before Their Appointments, and a Supervising Doctor Never Had to Step In

🌏 Global | Patient-Facing Tools

Google's AMIE is a conversational AI built to take a medical history and work up possible diagnoses. In this feasibility study with Beth Israel Deaconess Medical Center, patients with a single complaint talked to it from home up to five days before an urgent care appointment. Enrolment ran from April to November 2025, and 98 patients completed both the AI conversation and the appointment.

Supervising physicians watched each conversation in real time and could step in. Under the safety criteria set in advance, none of the conversations had to be interrupted. The design was single arm and single centre, funded by Alphabet, and the authors say larger trials are needed before anyone treats the result as proven.

Why it matters in a clinic: the tool did the intake before the clinician started. That is the job a long new-client form does in a physio, OT or speech pathology practice, and it is where vendors are likely to pitch AI next. The evidence so far says it can be done with a doctor watching. It does not say it can be done safely without one, and this is research, not a product you can buy.

What it means for clinicians:

• Look at your own intake form and first-session routine, and note the questions every client gets asked more than once. Those are where an AI intake tool would be pitched to you, and where you can judge a vendor's claim.
• If a vendor offers AI intake, ask who supervises the conversation, what happens when a client describes a crisis or a risk, and where the transcript is stored. The study had a doctor watching every conversation, so ask what your version has.
• Whatever a tool collects, confirm the history with the client in the room before it goes in the record. Treat any AI summary as a prompt for your questions, not a replacement for them.

A 20 Second Voice Clip Could Flag Type 2 Diabetes, in Work Co-Developed at a Melbourne University

🌏 Global | Screening

The model was trained on 63,283 voice samples from 21,129 people in the UK and US who had reported whether they had diabetes. It looks for voice changes linked to type 2 diabetes, such as increased hoarseness and reduced breath control. It was then tested on remote recordings of people reading one of Aesop's fables.

In the first evaluation, of 7,319 UK adults, the model gave a higher risk score to people who reported type 2 diabetes 80% of the time. A second check used 801 participants who took an HbA1c test at home within three months of the recording. The team says the model performed well across ages and genders and less well in Black participants, probably because there were few of them. Recordings can be collected over a phone or an app.

The researchers describe it as a screening aid, and a voice note cannot confirm diabetes. It belongs in front of an allied health clinician because of reach. Diagnosis of type 2 diabetes often lags, and the clinicians who see people regularly, such as exercise physiologists, physios, OTs, podiatrists and dietitians, are often the first to notice something is off. There is no peer-reviewed paper yet and no product to buy.

What it means for clinicians:

• Keep asking about diabetes history, family history and risk factors at intake, and record the answers. If a client mentions a voice or app screening result, treat it as a reason to see their GP for an HbA1c, never as a diagnosis.
• If you run exercise programs for clients with type 2 diabetes or prediabetes, record medication, glucose monitoring and hypoglycaemia screening in the file at each session. Wider screening, if it arrives, would bring more of these clients through your door.
• Do not tell clients this is available. It is a conference result, so file it as direction of travel.

Autonomous AI Triaged 94% of Urgent Skin Cancer Referrals at Two UK Hospitals

🌏 Global | Triage and Access

The system classified lesions from smartphone clinical and dermoscopic photos. Lesions it judged benign were discharged without a clinician looking, and higher-risk cases went to a teledermatologist. The AI discharged 31% of patients at one site and 25% at the other without review, and teledermatologists discharged a further 24% and 25%.

The headline figure of 8,500 appointments is an estimate. The team converted 2,851 clinician hours into face-to-face appointments, assuming 20 minutes each. Specificity was 72.1%, and the six missed cases came to light through post-market surveillance. This is a single deployment presented at a conference, and the researchers say it needs replication in larger populations and other health systems.

Your waiting list is the same problem on a smaller scale. The researchers frame the tool as reallocating scarce specialist time rather than replacing dermatologists, and the design keeps a human on the higher-risk cases. For allied health the pattern to watch is who decides that someone does not need to be seen. A filter that clears a quarter of referrals unreviewed is a staffing decision, and the missed cases are its cost.

What it means for clinicians:

• If a client has a new or changing skin lesion, send them to their GP. No AI triage tool changes that, and this one still missed six cancers.
• If you are offered any tool that sorts referrals or waitlist enquiries, ask what it does with the cases it is unsure about and how it tracks the ones it got wrong. Post-market monitoring is how these misses were found.
• Write down who reviews what in your own intake, so no enquiry is dropped by an automated filter, whether it sits in your booking software or your inbox.

Signals

🌏 Global | Wearables: A study in npj Women's Health, published on 30 September, used Fitbit data from 134 women with chronic pelvic pain, mostly endometriosis, and 61 healthy controls, worn for up to 90 days. Personalised models built on about 10 days of data predicted activity an hour ahead and flagged upcoming 15 minute sedentary windows, and simple interpretable models matched deep learning. It is a proof of concept, but for pelvic health physios it points to prompts that arrive before a long sit rather than after, which is a pacing conversation you can have now. Read →

🇦🇺 Australia | AI Regulation: The TGA says it does not intend to publish product-specific findings and is still working with suppliers on the status of particular features. The Medical Software Industry Association backs that, arguing that naming developers before safety is assessed could be prejudicial, while Digital Rights Watch says scribes should be regulated like any other medical device. Until that changes, the only way to find out whether your tool is affected is to ask the vendor. Read →

🌏 Global | Clinical Skills: Healio reports a Lancet Gastroenterology and Hepatology study in which adenoma detection rates in colonoscopies done without AI fell from 28.4% in the three months before AI arrived to 22.4% in the three months after. One expert's projection, not a measured result, is about 18% more missed cancers. It is a single observational study in gastroenterology, but the lesson reaches any clinician leaning on AI suggestions: do your own assessment first, so the skill is still there on the day the tool is off. Read →

🌏 Global | Predictive Models: The five-year, milestone-based project spends its first three years building and validating a digital twin that models a patient's immune response, so clinicians could test a treatment virtually before using it. The team estimates it could cut ICU length of stay by at least 25% if it works. It is years from any ward, so for acute and ICU physios and OTs the relevance is direction of travel: predicted recovery trajectories may eventually shape discharge and rehab planning. Read →