IDx-DR, now marketed as LumineticsCore by Digital Diagnostics, was authorized to screen adults with diabetes for more-than-mild diabetic retinopathy without an eye specialist reading the photo first. A primary care team captures images on a specified retinal camera and the software returns a refer-or-rescreen result. It is not a complete eye examination. Pregnancy and several existing retinal conditions sit outside the labeled use.
LumineticsCore (IDx-DR) · Digital Diagnostics · Topcon NW400
The MASAI trial in Sweden randomly assigned screening mammograms to AI-supported reading or standard double reading without AI. Transpara scored each exam and routed lower scores to a single reader and the highest score to double reading, with detection marks available on higher-risk exams. This card quotes the prespecified interim safety analysis, not a claim that every screening program should switch.
Transpara 1.7.0 · ScreenPoint Medical
Several cleared products flag possible urgent findings on a chest radiograph, such as pneumothorax, and move that study up a reading worklist. Names often cited for this job include Lunit INSIGHT CXR, Annalise Enterprise CXR, and qXR. This card describes the shared workflow, not a head-to-head, and it does not pool accuracy numbers across vendors. A flagged study still needs a radiologist’s report.
Lunit INSIGHT CXR · Annalise Enterprise CXR · qXR (Qure.ai)
The Targeted Real-time Early Warning System (TREWS) watches the electronic record and asks a clinician to confirm or dismiss a possible sepsis alert. It was studied prospectively across five hospitals, not as a silent background score. The published association depends on a person acting on the alert within a few hours. It does not diagnose sepsis by itself.
TREWS · Bayesian Health
The Epic Sepsis Model is a proprietary score switched on in a large number of U.S. hospitals. Researchers at Michigan Medicine compared it with sepsis in their own hospitalizations, instead of relying on the vendor performance sheet. The paper is included here because a familiar model can still fail local validation. Widespread deployment is not the same thing as demonstrated usefulness.
Epic Sepsis Model
This is an illustrative composite, not a trial report. A unit feeds monitor alarms through a second layer that withholds repeats already explained by motion, a loose lead, or a value the nurse just acknowledged. The bedside alarm for a new life-threatening rhythm still sounds. The point of the scenario is the governance question: who is allowed to silence what, and how a missed true alarm would be reviewed.
Multiparameter monitor · Local rules engine
This is an illustrative composite, not a trial report. Each morning the rapid-response nurse opens a list that combines a NEWS2-style vital-sign score with a model score from the record. The huddle decides who needs a bedside review. The score does not page a code team, order labs, or write a note. Hospitals using a vendor deterioration index should repeat a local check before trusting the threshold, as the sepsis literature already shows.
EHR deterioration score · NEWS2
Ambient systems record a consented visit and draft a note the clinician edits before signing. Abridge, Nuance DAX Copilot, and Nabla are three products sold for this job. Vendors describe less after-hours charting. Independent results differ by specialty, and a draft is not a signed note. Patients need to know the visit is being captured.
Abridge · Nuance DAX Copilot · Nabla
Computer-assisted coding suggests diagnosis and procedure codes from the note and sends them to a professional coder. Products in the 3M lineage, now under Solventum, have done a version of this for years, and newer language models are being added on top. The coder remains the person who submits the claim. A suggestion is not documentation of a condition the clinician did not record.
Solventum computer-assisted coding · 3M CAC heritage
Viz.ai ContaCT analyzes brain CT angiography and can text a neurovascular specialist when it suspects a large-vessel occlusion, in parallel with the usual read. FDA permitted marketing in 2018 as notification software. The specialist still reviews full images on a diagnostic workstation. The product is not a diagnosis and not a treatment decision.
Viz.ai ContaCT
This is an illustrative composite, not a trial report. As the triage nurse types the chief complaint and first vitals, a model suggests an acuity level. The nurse accepts or overrides it, and the accepted level is the one that enters the queue. The scenario is about keeping a human on the decision when under-triage would delay care. No accuracy percentage is claimed, because this entry is not a study.
EHR triage form · Local acuity model
Paige Prostate marks a location on a digitized prostate biopsy slide that the model ranks as the most suspicious for cancer, after the pathologist’s own first look. FDA authorized it in 2021 as an adjunct. The output is not the diagnosis. The pathologist still signs out, using the full clinical context and any extra stains they choose to order.
Paige Prostate · Paige FullFocus
Vendors including PathAI and Ibex offer algorithms that estimate staining extent, such as a Ki-67 or PD-L1 percentage, on a digitized slide. A pathologist checks the region and the number before it enters the report. This card does not restate concordance percentages, which are assay-specific and easy to over-read. The signed score remains a pathology result, not a model result.
PathAI · Ibex
Insilico Medicine used a generative platform to propose TNIK as a target in idiopathic pulmonary fibrosis and to design rentosertib (formerly ISM001-055). Nature Medicine published a 12-week randomized phase 2a trial in 2025. The primary endpoint was safety, not lung function. The molecule is investigational. A small secondary signal is not an approved therapy.
Pharma.AI · Insilico Medicine · rentosertib (ISM001-055)
Trial-matching tools read structured fields and sometimes the free-text note to list studies a patient might qualify for. Tempus, Flatiron, and Deep 6 AI are examples of companies that sell a version of this. A research coordinator still checks every criterion against the chart. A match list is not enrollment, and it is not a treatment recommendation.
Tempus · Flatiron · Deep 6 AI
This is an illustrative composite, not a trial report. Daily weight, blood pressure, and a short symptom check land in a pool. A simple model ranks who should be called today when the heart-failure nurse starts the shift. The nurse places the call and decides on a diuretic plan with the covering clinician. The scenario deliberately stops before any claim that admissions fell. Device trials of hemodynamic monitors are a different evidence base and are not borrowed here.
Home weight and blood pressure · Clinic message pool
This is an illustrative composite, not a trial report. Before a diabetes visit, software drafts a short note from a continuous glucose export: time in range, overnight lows, and the days that look different. The clinician edits it and decides any change in insulin. The scenario excludes autonomous dosing. Closed-loop insulin systems are regulated devices with their own labels and are not this workflow.
Continuous glucose monitor export · Clinic inbox
This is an illustrative composite, not a trial report. At discharge, a model drafts a short explanation of the diagnosis, medicines, and warning signs from the clinician’s own note, in the patient’s language where the hospital supports it. The clinician or pharmacist edits every line before it is printed or sent. The draft is not allowed to add a new instruction that is absent from the signed note. It is not a chatbot the patient can ask for new medical advice.
EHR discharge note · Clinician edit step
Ada is a symptom-assessment app that interviews a person and suggests possible causes plus a level of urgency. It is a patient-facing product, not a clinician order. Regulatory status and intended use differ by country. An assessment is not a diagnosis, not a prescription, and not a reason to delay emergency care when red-flag symptoms are present.
Ada
Wysa is an example of a conversational program that offers cognitive-behavioral exercises between visits with a person. It is support, not a therapist, and it is not a crisis service. Someone at immediate risk needs local emergency care or a crisis line, not a chatbot. Health systems that offer a tool like this still need a path to a human.
Wysa
REACH VET is a U.S. Veterans Health Administration program. A statistical model flags veterans in a very small top tier of risk each month, and a clinician who knows the patient reviews the chart and makes contact. Identification is not a diagnosis of suicidal intent. The program exists to prompt a human conversation and a look at whether care should change.
REACH VET · VHA electronic health record
This is an illustrative composite, not a trial report. Late in the afternoon a model estimates how heavy the next shift will be from the current census, planned admissions, and recent acuity. The charge nurse uses it as one input to ask the staffing office for help. It does not cut a shift by itself, and it does not replace the legal and contractual staffing rules of the hospital. No savings percentage is stated.
Staffing office workbook · Ward census
This is an illustrative composite, not a trial report. The rounding list shows who is at higher fall risk and which precautions are already ordered, so the nurse does not hunt through the chart in the doorway. It is a prompt, not a new score the nurse must obey. Alert fatigue is part of the design problem: if every patient is flagged, the list has failed. No fall-rate reduction is claimed.
EHR fall-risk score · Bedside rounding list
This is an illustrative composite, not a trial report. For a planned imaging study or procedure, a model assembles the indication, prior treatments, and relevant results into a draft the authorization staff edit before anything is sent to a payer. Staff have to check that each sentence is actually in the chart. Models omit and sometimes invent. No approval-rate improvement is claimed.
EHR chart · Payer portal
This is an illustrative composite, not a trial report. After a payer denies a claim, a model drafts an appeal that quotes the denial reason and the passages in the note that speak to it. A coder or clinician edits the letter and decides whether the appeal is worth sending. The draft cannot create a diagnosis that was never documented. No overturn rate is claimed.
Denial letter · Clinical note
EndoNet, from Twinanda and colleagues, is a research network that recognizes surgical phases and instruments in laparoscopic cholecystectomy video. It was demonstrated on the public Cholec80 collection. The result is about whether a model can tell where a case is in its sequence. It is not evidence that video analysis reduces complications, and it is not a product cleared to guide a surgeon’s hands.
EndoNet · Cholec80
Robotic systems such as the da Vinci platform can store console time, instrument use, and in some setups video that a surgeon reviews after the case. Vendors present this as coaching and quality review. The surgeon still operates. Nothing in this card says that analytics lower complications. That would be a separate study of a defined program.
da Vinci robotic platform · Post-case reports
The U.S. CDC National Wastewater Surveillance System publishes pathogen trends from sewage, which health departments read next to emergency and lab data. Some jurisdictions add a statistical nowcast on top of that feed. This card is about the desk workflow, not a claim that a named model forecasts incidence with a stated error. Wastewater is a population signal. It does not diagnose a person.
CDC NWSS · Local epidemiology desk
This is an illustrative composite, not a trial report. A registry ranks patients who are overdue for a fit test, mammogram, or other screening the health system already recommends. Navigators call down the list. Before anyone trusts the rank order, the team checks whether the score under-calls a language group, a neighborhood, or a payer. No detection-rate improvement is claimed. A biased rank can widen a gap the program meant to close.
Population health registry · Outreach workqueue