The AI Health Index Brief

The week in one line

Two prostate cancer tests stopped estimating risk and started choosing treatment, and an inpatient coding platform said it no longer needs a human on the case. Four FDA decisions added new readings to machines hospitals already own, which makes them the rare upgrade that never needs the loading dock.

This issue covers two weeks, September 27 to October 9, plus three earlier entries that arrived after the last issue went out, one of them an FDA authorization from July. That makes 32 entries across 30 vendors, 27 Verified at source and 5 Partially Verified. Product changes led with 13 entries, and regulatory decisions, clinical evidence and EHR integration followed with five each.

The test started choosing the treatment

Two prostate cancer tests brought new evidence to ASTRO 2026, the annual radiation oncology meeting, and both made the same move. They stopped only estimating how dangerous a cancer is and started saying which treatment will help.

Veracyte reported an analysis of NRG/RTOG 0815, a randomized phase III trial, for its Decipher Prostate test. The trial asked which men with intermediate risk prostate cancer gain from adding short term hormone therapy to radiation, and it followed 395 patients for a median of 10.2 years.

Men with intermediate or high Decipher scores saw distant metastasis at ten years fall by 8 percentage points with hormone therapy. The 44 percent with low scores saw a 1 point difference that was not statistically significant.

The second number is the valuable one. Hormone therapy has real side effects, and a test that identifies who gains little from it can spare those men the treatment. A randomized trial, ten years of follow up and a treatment decision at the end is the evidence guideline panels read first.

ArteraAI announced five ASTRO presentations on its multimodal AI prostate biomarker. One is the first validation in a South Asian population. In the POP-RT trial in India, patients it scored as high risk had fewer distant metastases with whole pelvis radiation, while intermediate risk patients saw no similar benefit.

Another compared risk estimates from commercial testing with STAR-CAP, a database of nearly 20,000 patients, and found them in line with observed outcomes in every NCCN risk group. A third found no evidence of bias by race or age in NRG/RTOG post prostatectomy trials. A model checked against a new population, a large real world database and its own subgroups is a model that expected the questions.

Caris pushed its liquid biopsy the same way. The enhanced Caris Assure combines whole exome, whole transcriptome and whole blood analysis in one assay, so changes coming from the tumor can be told apart from clonal hematopoiesis and incidental germline findings.

For patients tested repeatedly during treatment, results now come back sorted as molecular response, intermediate or molecular progression, in a median of about seven days. That turns a one time profile into a running answer to whether the treatment is working. Caris says tumor fraction tracks progression risk, but the release does not publish the validation behind it.

Brainomix reported new analyses from INBUILD, a phase III trial in progressive pulmonary fibrosis. In 474 patients from the CT substudy, its e-Lung measures detected the effect of nintedanib, and greater disease extent at baseline went with faster decline in lung function over 52 weeks.

The e-Lung results broadly agreed with a UCLA research algorithm run on the same scans. It is a post hoc analysis, so it makes the case for quantitative CT as a trial endpoint before it makes the case at the bedside.

Our read

For most of this log’s life, the typical evidence claim has been accuracy, and the better ones have been about risk. This period the strongest entries answered a harder question: who should get the treatment, and is it working. A prognostic score informs a conversation. A predictive one changes an order, and orders are what payers reimburse and guidelines write down. Every diagnostic in this index will eventually be asked to make that claim.

The hospital ran the study

Two more evidence entries came from the customer side. AZmed’s fracture detection has run on every patient at CHU de Reims since 2021, and the hospital published its own evaluation in Pediatric Radiology.

In 366 children seen in the pediatric emergency department for limb trauma, Rayvolve reached 93.6 percent specificity. With the AI, the less experienced resident’s sensitivity rose from 87.0 to 91.3 percent and specificity from 88.2 to 93.2 percent.

Specificity is the right number to lead with in children. A false alarm means immobilization, a follow up visit and a worried parent for a fracture that was never there. The study is single site with two readers, so it adds to larger validations rather than replacing them.

Primary Partners Alliance, a physician owned network in Central Florida that runs on Lightbeam, reported a year of results across its MSSP and ACO REACH patients. Patients who completed an annual wellness visit cost nearly $70 less per member per month, $947 against $1,014.

They were readmitted 9 percent of the time against 16 percent, and their HCC recapture was 70 percent against 57. Named numbers from a named network are rarer than they should be in population health.

The comparison is between patients who did and did not complete a visit, and people who come in for a wellness visit are not a random sample. The figures measure the engagement program as a whole. They cannot separate the software from who walked through the door.

Our read

Evidence a customer publishes reads differently from evidence a vendor publishes, because the vendor cannot shape it. That makes it more credible and harder to plan for. The best of it tends to come from sites like Reims, which have run a product long enough to stop being impressed by it.

Autonomy came with a dial

AKASA launched an autonomous platform for inpatient coding and clinical documentation integrity, building on its prebill review products. The company says it fully codes complex inpatient cases with no human intervention, in under 90 seconds after discharge, using models tuned to each health system.

Inpatient coding is where autonomy has been hardest to trust. A wrong code carries reimbursement risk and compliance risk at once, and a hospital stay produces a long, messy record. AKASA reports that blinded third party evaluations found its coding matched or exceeded human coders on MS-DRG assignment, principal diagnosis and present on admission accuracy.

Those evaluations are not published. AKASA says it designs a phased rollout with each customer, and that rollout is where a health system finds out how much of the full autonomy claim holds in its own records.

Arcadia launched a new portfolio at its Aggregate conference, splitting its products into lines for providers and payers, for market and network intelligence, and for real world evidence. Much of it renames capability Arcadia already had. The new parts are a care navigation agent, AI patient prioritization and a stated oversight model.

The stated model is the interesting part. Arcadia says its AI runs on a spectrum from assistive to autonomous depending on risk, with a full record of what it did and why. A spectrum is the honest shape for this. Where each workflow sits on it today is what the launch leaves open.

Upheal added Routines, which let its assistant run scheduled background tasks. Built in routines cover a daily billing wrap up and session prep briefings, and clinicians can write their own in plain language, such as a check for missed sessions or denied claims. The assistant drafts the action and the clinician approves it.

An assistant that works on a schedule has changed jobs. It used to wait to be asked. Now it does the rounds, and it still needs a signature before anything leaves the building.

Enzo Health shipped Clinician Verify for Enzo EHR customers. Every OASIS suggestion from QA or a clinical manager now goes to the clinician who wrote the documentation. On their phone they see the original answer, the suggested one and the reasoning, then choose which answer stays before they sign.

That reverses the usual flow, where a reviewer fixes the assessment and the author may never learn why. In home health, OASIS answers drive payment, so a recorded trail of who changed what is worth having on its own.

Our read

Autonomy claims have started arriving with a setting attached. AKASA promises no human on the case and a phased path to get there. Arcadia describes a spectrum, Upheal drafts and waits, and Enzo routes the correction back to its author. The product question has moved from whether the AI acts alone to where the dial sits for each workflow, who can move it, and what the record shows afterward. The Index Answer below counts the vendors that write that down.

New readings from scans already taken

Tempus received FDA 510(k) clearance for ECG-MR, which reads a standard resting 12 lead ECG and flags signs of moderate or severe mitral regurgitation. It is cleared for adults 65 and older with cardiovascular risk factors and no prior mitral valve diagnosis or intervention.

The result is binary and meant to prompt a referral for imaging, not to diagnose. It is the fourth cleared Tempus ECG product, following pulmonary hypertension in August. Mitral regurgitation is often found late because early disease has few symptoms, and this screen runs on ECGs that are already being taken.

AIRS Medical received 510(k) clearance for SwiftSight Body Composition. From an MRI scan of under five minutes, it measures fat, muscle and liver fat and benchmarks each against age and sex matched reference populations in one report. SimonMed Imaging is the first customer.

The clearance takes SwiftSight beyond brain imaging, with diffusion tensor imaging planned for November. Because MRI uses no radiation, scanning the same patient again to track change over time is practical.

Caristo Diagnostics holds a De Novo authorization, granted in late July, for CaRi-Heart. It quantifies coronary inflammation from routine coronary CT angiography and estimates ten year cardiovascular mortality risk. Alongside CaRi-Plaque, cleared in February 2025, plaque and inflammation can now be read from one scan.

Caristo planned a US commercial launch starting in the third quarter. It has not yet published the US indication or any conditions attached to the authorization, and those details decide how the risk estimate can be used in care.

Ceribell received 510(k) clearance for quantitative EEG trending on its point of care system, including alpha delta ratio and relative alpha trending, burst suppression monitoring and amplitude integrated EEG. The clearance carries no age restriction, so the same trends apply from premature newborns in the NICU to adults in the ICU.

The features arrive by automatic software update in the first half of 2027. Cleared now and shipping later is becoming a familiar pattern, and it turns a clearance into a dated line on a product roadmap.

Our read

The hardware in these four entries was already in the building. The ECG cart, the MRI scanner, the CT scanner and the EEG system were bought and billed long before the FDA got involved. Each time, what was authorized is a new reading of a signal the hospital was already collecting. For product teams that changes the pitch, from a new device to more value from a scan the customer has already paid for.

The AI moved into software already on screen

Several entries this period were about where the AI lives rather than what it does, and three of them were built for nurses. voize’s voice documentation for nurses now writes directly into ORBIS U, the hospital information system from Dedalus HealthCare.

Nurses speak observations at the bedside, voize structures them, and the entries land in the ORBIS record where the rest of the team can use them. The company says Agaplesion runs it in production and that, after a three month pilot, it saves up to an hour of documentation per nurse per shift. That figure comes from one customer, so it is a target to test rather than a benchmark.

Caregility and Ascom finished an integration that lets bedside staff call in a virtual nurse from the Ascom nurse call controls already in the room. One button press connects the team to a remote clinician through the Caregility smart room device. Virtual nursing programs stall when staff have to leave their workflow to ask for help, and this removes the detour.

Elsevier extended ClinicalKey AI to nurses with ClinicalKey Nursing AI, a conversational decision support tool that answers clinical questions at the point of care and shows the evidence behind each answer. Most clinical assistants were designed around physician questions. Nurses ask different ones.

The pattern held outside nursing. Corti’s ambient documentation is now built into PATFAK 6.6, Redline Data’s software for German rehabilitation, counseling and social care facilities. Therapy sessions recorded in the PLINK app become progress notes, and the reporting module drafts care plans and applications from treatment records.

Roughly 1,500 facilities run PATFAK, and they get AI documentation without bringing in another vendor. Use is billed in credits with a monthly free quota for each seat, a pricing model built for trying the thing before anyone writes a budget memo.

Chartnote reached the same place by a different road. Version 3.3.7 of its Chrome extension adds Companion, which pins a Chartnote note beside any web based EHR so clinicians can insert sections into the chart with @section shortcuts. There is no integration project, because the browser is the integration.

Riverain Technologies made its ClearRead chest imaging AI available through RapidAI’s enterprise platform and radiology workflow. Spring Health embedded its platform in Workday Wellness, one click from where employees already go. Lightbeam connected ChartSpan’s remote monitoring data to the population health workflows its care teams already use for risk and outreach.

Our read

Distribution is turning into the product. Each of these vendors picked a host the customer already opens every day, from the hospital information system to the nurse call panel to the browser. Last issue a trials platform shipped its agent inside a general assistant. This period it was the hospital’s own furniture. The integration announcement increasingly carries more news than the model announcement.

The small fixes carried the safety

Some of the most consequential changes this period took one line in the release notes. Artera, the patient communications platform, now sends blocked numbers an automated emergency response telling them to call 911 or reach their care team by text or spoken message. A patient whose number has been blocked can still be told where to turn.

NeuroFlow now sends urgent alert emails the same day instead of the next. In behavioral health, the time between a patient flagging risk and the care team hearing about it is the gap the product exists to close, and this release closed most of a day of it.

Syllable shipped two releases two days apart. The first fixed an edge case that caused duplicate outbound campaign calls, which reach patients directly. It also added GPT-6 Sol and GPT-6 Luna to the OpenAI models its voice agents can use.

The second added five Gemini 3.x models and switched off the temperature setting for the two new GPT-6 models to match gateway support. It also added deprecation warnings for older language models and for Google speech to text V1. That is seven new models in two releases, plus a warning system for the ones on their way out.

Our read

The model menu inside healthcare platforms now changes faster than most procurement reviews can convene. Deprecation warnings are the right instinct, because the bigger risk is rarely a bad new model. It is a good old one retired underneath a live workflow. The safety work that reached patients this period was quiet: an emergency message for a blocked number, an alert sent today instead of tomorrow, one call instead of two. None of it would survive a keynote. All of it matters more than most keynotes.

Market notes

Xaira Therapeutics reported results from the Vega class of its X-Design models, which design antibodies with no data from earlier campaigns. Against its lead oncology target, one design pass produced a binder worth progressing in three weeks and a preclinical lead in seven.

Against a GPCR target where library screening and llama immunization had both failed, a library of 60,000 designs yielded a developable antagonist now in lead optimization. Beating a target that defeated the llamas is the claim that matters, and Xaira says a systematic assessment of X-Design will follow.

IQVIA launched Life Science Models, a suite built on NVIDIA Nemotron and AWS that links molecular and treatment information with patient journeys. The stated uses include study simulation, protocol optimization, digital twins and synthetic control arms. Synthetic controls still face close regulatory scrutiny, so the near term value is in better protocols rather than replaced trial arms.

GeneDx introduced Easy Order at the American Academy of Pediatrics meeting, an exome and genome workflow built for general pediatricians rather than genetics specialists. Testing in children often waits in referral queues for genetics clinics, and this is a bid to move first line testing into routine care.

Wisedocs launched Legal Intelligence in early access for insurance defense teams. It checks deposition transcripts and demand letters against a claim’s medical records, and every flagged contradiction cites both the testimony and the record, which is what makes a flag usable in front of a lawyer.

CarePilot’s September 29 release lets a practice set note structure once for the whole organization and lets clinicians undo a full note pull forward. Billing suggestions for time based codes now follow current Medicare guidance.

Alcidion’s Miya Precision Concept Detection, which reads free text notes and suggests SNOMED CT codes for a clinician to review, is now registered as a Class I medical device in Australia and with the UK’s MHRA. Class I is the lowest risk class, so the registration speaks to process more than performance. It does clear a bar that hospitals in both countries increasingly set for AI in documentation.

What the two weeks say about the category

Thirty two entries, and the theme is consequence. The prostate tests moved from estimating risk to choosing treatment. AKASA moved from reviewing inpatient codes to assigning them. The FDA decisions added readings to machines hospitals already own, and the integrations put AI inside software people already have open.

Each move raises the bar in a different place. When a test picks the treatment, its evidence has to be predictive rather than prognostic. When an agent codes the chart, its oversight model has to be written down. When the AI lives inside someone else’s software, the integration is the product.

The category is doing more, in more places, and it is being asked to show its work in all of them at once.

Index Answer

Which healthcare AI vendors publish what their AI may do without a human?

Few publish the whole answer, and many publish half of it. Of the 561 vendors the AI Health Index has assessed on Autonomy and Oversight Model as of October 5, 2026, 32 grade A. An A requires both halves in writing: what the system may do and what it may not, with the escalation threshold, the override path and the conditions that send a case to a person.

Another 354 grade B. They describe an oversight structure with one part missing, most often the point at which the system stops or what happens after it gets something wrong. That missing part is usually the one that matters on a bad day.

The remaining 175, 106 at C and 69 at D, assert oversight without describing a mechanism or publish nothing at all. Human in the loop written as a phrase rather than as a described control grades C. A claim that the system does not make mistakes grades D, because a buyer who believes it will not build the review step that would catch one.

This period showed how varied the written answers can be. AKASA describes full autonomy on complex inpatient cases inside a phased rollout. Arcadia describes a spectrum from assistive to autonomous with a record of each action. Upheal drafts and waits for approval, and Enzo sends every correction back to the clinician who wrote the original. Those are four different models, and each is useful to a buyer only because it was published.

The axis grades disclosure, not caution. A highly autonomous product with a published threshold and override path can grade A. A cautious product that will not say what happens when it is unsure grades poorly, because an oversight model nobody can read looks the same from outside as one that does not exist.

Full grades, the axis definition and what separates each band are on the Autonomy and Oversight Model page.