Updated November 2026 · Written and maintained by the Progression Agency strategy team
AI diagnostics software is judged on one sentence before anyone looks at the model: the intended use. Whether the output moves a study up a reading list, measures something on an image or suggests a diagnosis decides who acts on it, what evidence it needs and which FDA path fits. FDA’s list of AI-enabled medical devices held 1,614 entries at its September 4, 2026 update, and 1,230 of them name radiology as the lead panel. This page follows a diagnostic model from that sentence through data, labels, validation, regulatory files, deployment and monitoring. Progression’s own developers build the software, and the team is based in New York.
On this page · 14 sections
- Two sentences, one chest X-ray model
- AI diagnostics starts with the intended use sentence
- Where AI in diagnostics is already authorized
- Data for AI medical diagnosis models
- Labels and the reference standard
- Validating AI in diagnostics before anyone relies on it
- The FDA path for diagnostic software
- Change control plans for model updates
- Generative AI and healthcare ChatGPT tools near the diagnosis
- Fitting the model into the reading workflow
- Monitoring after launch
- What AI diagnostics software costs to build
- How Progression’s developers build diagnostic AI
- Before the first line of code
Building health software without a model? Patient portals, EHR connections and clinical workflow tools that carry no diagnostic model are covered on our healthcare software development page.
Examples marked “Illustrative example.” are hypothetical; no real patient, device, hospital or company is shown.
Two sentences, one chest X-ray model
Picture a hypothetical team with a model that looks at adult chest X-rays. Described one way, it flags studies with a suspected finding so a radiologist reads them sooner. Described another way, it tells a patient whether the image shows pneumonia.
The weights are identical, yet the two descriptions make two different products. The first keeps a specialist between the software and every result; the second puts the software’s answer in front of a person with no one to interpret it. The user, the evidence, the labeling and the regulatory questions all change with one sentence, which is why it gets written first.
- Write the intended use before choosing data.
- Name the first person who sees the output.
- Name the setting: emergency department, outpatient clinic, home.
- Let every later stage answer to that sentence.
AI diagnostics starts with the intended use sentence
Most diagnostic models fall into a handful of jobs. The table sets each job beside example wording, the person who acts on the output and the questions the sentence raises for the build.
| What the output does | Example wording | Who acts on it | What the sentence raises |
|---|---|---|---|
| Prioritize | Moves suspected cases up a reading list | The specialist who reads every study | Time to read, missed flags, the order of the list |
| Detect and mark | Draws a box or a heat map on a possible finding | The reader, who accepts or dismisses each mark | False marks per study, marks a reader ignores |
| Measure | Reports a volume, a length or a ratio | The clinician who signs the report | Agreement with manual measurement across scanners |
| Suggest a diagnosis | Names a likely condition with a score | The clinician making the diagnosis | Evidence for each condition named, and subgroup results |
| Give a patient a result | Tells a person what a test shows | The patient, possibly alone | Readability, follow up instructions, a much higher bar |
| Draft text for a clinician | Writes a first draft of a report or summary | The clinician who edits and signs it | Errors of omission, invented findings, review time |
Who reads the output first?
A specialist who reads every study can catch a wrong flag, so a triage tool’s errors cost time rather than a missed diagnosis. A patient reading a result alone has no such check, so a sentence that reaches patients needs far more evidence, plainer words and clear next steps.
Words that widen a claim
Small verbs change what a team must prove. Keep them out of the sentence, the screens and the sales copy unless the evidence covers them.
- “Diagnoses” where the software actually flags.
- “Rules out” where the test only measured detection.
- “Replaces” where a reader still signs.
- “All” or “every” in front of a finding.
- A setting or age range the test data never covered.
Where AI in diagnostics is already authorized
FDA keeps a list of AI-enabled medical devices authorized for marketing in the United States. Each device on it met the applicable premarket requirements, and each entry links to the database record with its public summary. At the September 4, 2026 update the list held 1,614 entries, with final decisions running from September 1995 to June 29, 2026.
By submission number, 1,553 entries begin with K, the 510(k) prefix; 40 begin with DEN for De Novo; and 21 begin with P for premarket approval. Radiology is the lead panel for 1,230 entries, cardiovascular for 154 and neurology for 73.
What does the list leave out?
FDA says the list is not comprehensive. Devices were identified mainly from AI terms in the summary descriptions of their marketing authorizations and from their classification, and devices whose decision summaries were not yet published wait for a later update. FDA also says it will explore ways to tag devices built on foundation models, from large language models to multimodal designs.
Why imaging leads the count
Images are already digital, stored in standard formats and read by specialists who can check a model’s output. That makes imaging a natural first home for diagnostic software, and it means a new imaging model meets an established set of products and reviewers’ expectations.
- Search the list by product code to find devices with a similar intended use.
- Read the public summaries for how others described their testing.
- Note which entries mention a change control plan.
- Check the update date before quoting any count.
Data for AI medical diagnosis models
Training data usually comes from a hospital’s imaging archive, lab system or EHR, or from other hospitals under a data use agreement. Before it reaches a training run, the data is de-identified under HIPAA’s standard in 45 CFR 164.514, which offers two routes.
The first is expert determination: a person with appropriate statistical and scientific knowledge finds the risk of identifying anyone very small and documents the methods and results. The second removes 18 listed identifiers of the individual and of relatives, employers or household members, and the covered entity has no actual knowledge that what remains could identify someone. A re-identification code is allowed if it is not derived from the person’s information and its mechanism is not disclosed.
| Identifier type in the rule | Where it hides in diagnostic data | What the pipeline does |
|---|---|---|
| Names | Image headers, report text, scanned requisitions | Strip or replace with a study code |
| Dates except the year | Study, birth, admission and discharge dates in headers and reports | Keep the year or shift dates consistently by patient |
| Medical record and account numbers | Header fields and report footers | Replace with a code not derived from the number |
| Device identifiers and serial numbers | Scanner serials in image headers | Remove, keeping only the scanner model where needed |
| Full face photographs and comparable images | Clinical photos, some head and neck studies | Exclude, crop or review case by case |
| Geographic subdivisions smaller than a state | Addresses and ZIP codes in reports | Remove, with the rule’s three digit ZIP exception only where it applies |
| Any other unique code | Text burned into pixels, free text notes | Scan pixels and text, and remove what is found |
Pixels and headers carry identifiers
Image files carry text fields with names, dates, record numbers and scanner serials, and some images have text burned into the pixels themselves. A pipeline that cleans the fields but never scans the pixels can leave a name on the image. Head and neck studies need a separate look under the rule’s line on full face photographs and comparable images.
Split by patient, then by site
Split training, tuning and test data by patient, never by image, so no person appears on both sides. Then hold one or more sites out entirely, so the final test shows how the model behaves on scanners, protocols and populations it has never met.
- Record which archive and date range each study came from.
- Log the de-identification method used for each batch.
- Keep the code key, if any, away from the training environment.
- Freeze the test set before model work begins.
Labels and the reference standard
A model learns whatever the labels say, so the definition of the right answer is a design decision. The reference standard might be a specialist read, a consensus of several readers, a pathology result or a later clinical outcome, and each choice costs different time and money.
Write the labeling guide first
The guide says what counts as present, absent and uncertain, how to mark location, and what to do with poor quality studies. Readers label a pilot batch, the team compares their answers, and the guide is revised before the full set is labeled.
When two readers disagree
Disagreement is information. Send split cases to an adjudicator, record both original answers, and track agreement rates by reader and by finding, because a finding readers cannot agree on is one a model cannot learn cleanly.
- Define present, absent and uncertain in writing.
- Pilot the guide on a small batch.
- Adjudicate disagreements and keep the original reads.
- Track agreement by finding.
Validating AI in diagnostics before anyone relies on it
Validation answers whether the intended use sentence holds on data the model never saw. Each measure answers a different question, and each can mislead on its own, so a report shows several together with the counts behind them.
| Measure | The question it answers | Where it misleads |
|---|---|---|
| Sensitivity | Of the cases that have the finding, how many did the model catch? | Looks high when the hard cases were left out of the test set |
| Specificity | Of the cases without it, how many did the model leave alone? | Hides the daily burden when the finding is rare |
| Positive predictive value | When the model flags a case, how often is it right? | Changes with prevalence, so it shifts from site to site |
| Negative predictive value | When the model stays quiet, how often is that right? | Can look reassuring simply because the finding is rare |
| Area under the ROC curve | How well scores separate positive from negative cases overall | Says nothing about the threshold the product will actually use |
| Calibration | Does a score of 0.8 mean right about 80 percent of the time? | Ignored when only a yes or no is shown, then trusted anyway |
| Time to read | Did flagged studies reach a reader sooner? | Measured in a lab instead of the real worklist |
Test sets from sites the model never saw
A model tested only on data from its training hospitals carries that hospital’s scanners, protocols and patient mix into its score. Hold out whole sites, report results per site and per scanner model, and treat a large gap between sites as a finding to explain before launch.
Does the nondiscrimination rule reach decision support tools?
It reaches covered entities that use them. Under 45 CFR 92.210 a covered entity must not discriminate on the basis of race, color, national origin, sex, age or disability through patient care decision support tools, has an ongoing duty to make reasonable efforts to identify tools whose input variables measure those traits, and must make reasonable efforts to mitigate the risk. Subgroup results in the validation report give a hospital what it needs for that review.
- Report results by age band and sex at minimum.
- List every input variable the model uses.
- Flag any input that measures a listed trait.
- Repeat the subgroup report after each update.
Have a model idea and a data archive?Tell us the intended use sentence you have in mind and where the data sits. You get a free proposal for the build, the records and the price.
Get a free proposalTell us about a project you’d like help with.
The FDA path for diagnostic software
FDA uses the IMDRF definition of Software as a Medical Device: software intended for one or more medical purposes that performs them without being part of a hardware medical device. It is one of three types of device related software, beside software in a medical device and software used to make or maintain one.
| FDA submission | How FDA describes it | When a diagnostic model team looks at it |
|---|---|---|
| Premarket Notification 510(k) | For Class I and II devices that are not exempt from premarket review or subject to another submission | A predicate device with the same intended use already exists |
| De Novo Classification Request | For novel low to moderate risk devices with no predicate device | The model is new and the risk is low to moderate |
| Premarket Approval (PMA) | For Class III devices | The output carries the highest risk |
| Investigational Device Exemption (IDE) | For investigational devices used in a clinical study to collect safety and effectiveness data | A clinical study is needed before a submission |
| Predetermined change control plan | Reviewed inside a 510(k), De Novo or PMA for an AI-enabled device | Updates after launch are already foreseeable |
FDA’s premarket submissions page describes each route. For AI-enabled devices, FDA has also published a draft guidance from January 2025 on lifecycle management and marketing submission recommendations, which it labels not for implementation, and its Good Machine Learning Practice page points to IMDRF’s final document of January 2025 with 10 guiding principles, building on principles FDA, Health Canada and the UK’s MHRA released in October 2021. The wider device software process sits on our medical device software development page.
Which submission fits a model with no predicate?
FDA describes the De Novo request as the route for novel low to moderate risk devices with no predicate device, while the 510(k) serves Class I and II devices that are not exempt. The intended use sentence and the risk of the output drive that choice, and the manufacturer’s regulatory team makes it.
Change control plans for model updates
Models improve with new data, and each change used to raise the question of a new submission. FDA’s final guidance of August 2025 on predetermined change control plans for AI-enabled device software functions gives another route.
A plan describes the planned modifications, the method to develop, validate and put them in place, and an assessment of their impact. FDA reviews it as part of the marketing submission, so modifications the plan describes can be implemented without a new submission for each one. The guidance applies to AI-enabled devices reviewed through the 510(k), De Novo and PMA pathways.
What belongs in the modification list?
Changes the team can already foresee: retraining on data from new sites, a new scanner model, a threshold adjustment, or a performance improvement measured the same way as the original. Changes to the intended use itself sit outside that kind of list.
- Retraining on new sites, with the same test method.
- Support for an added scanner model.
- Threshold tuning inside stated limits.
- A versioned record of every release.
Generative AI and healthcare ChatGPT tools near the diagnosis
The applications of generative AI in healthcare closest to diagnosis are drafting: a first version of a report, a summary of a long history before a visit, a patient message that explains a result in plain words. In each, a clinician reads, edits and signs.
Where a chat assistant stops
A general chat assistant was not built or tested for a diagnostic intended use, so it drafts and summarizes while the diagnosis stays with the clinician and any regulated software. Keep patient data out of tools that lack the agreements HIPAA requires; our HIPAA compliant software development page covers model provider agreements.
- Drafts carry a visible label until a clinician signs.
- Summaries link each statement to the record it came from.
- Patient messages are reviewed before sending.
- Every prompt and output is logged with the model version.
Assistants and chat features built for clinical teams are covered on our AI chatbot development page.
Fitting the model into the reading workflow
A diagnostic model earns its keep only where a clinician already looks. That might be the priority order of a worklist, an overlay in the image viewer, a field in a structured report or a note in the EHR. Each placement changes what the reader sees and when.
- Worklist priority for triage models.
- Viewer overlays for detection models.
- Structured report fields for measurement models.
- EHR notes for results a care team acts on.
- A clear way to dismiss or correct the output.
Run it silently first
Before go live, run the model on live studies without showing anyone the output, and compare its answers with the readers’ reports. Silent mode shows site specific gaps, integration bugs and timing problems while no patient depends on the result.
Mobile and patient facing pieces of the product are covered on our healthcare app development page, and image pipelines outside medicine on our computer vision development page.
Monitoring after launch
Scanners are replaced, protocols change and patient mix shifts, and the model meets all of it after launch. Monitoring keeps the evidence current. The reports and data feeds that track a model after launch are covered in our guide to healthcare data analytics.
- Override rate: how often readers dismiss the output.
- Flag rate by site and scanner model.
- Input checks for image types the model was not built for.
- Complaint and error reports routed to the product team.
- A version log tying each output to the model that produced it.
What should trigger a retraining review?
A sustained rise in overrides, a new scanner or protocol at a site, a shift in flag rate, or a subgroup result that moves. A review decides whether the change fits an existing change control plan or needs the regulatory team’s attention.
What AI diagnostics software costs to build
Diagnostic software is priced by the data work, the integrations and the records the product needs. The bands below run from the typical low to the published high in the US market for software builds; clinical studies and regulatory fees sit outside them.
| Build | Typical low to published high | Notes |
|---|---|---|
| AI proof of concept | $15,000 to $40,000 | One question tested on retrospective data |
| AI feature in existing software | $15,000 to $60,000 | A model added to a viewer, portal or EHR tool already in use |
| Integration or internal tool | $30,000 to $100,000 | Worklist, viewer or report connections |
| AI MVP | $40,000 to $120,000 | A first product with the model at its center |
| Platform or portal | $75,000 to $400,000 | Multi site, multi role software around one or more models |
| US agency or contract development rate | $100 to $200 an hour | Bill rates, not employee wages |
Progression’s price is set in the written proposal after scope and is subject to change. Our AI app development page shows how model features fit into a wider product budget.
What moves the price?
The number of sites and archives, how much labeling and adjudication the reference standard needs, the depth of integration with viewers and the EHR, and the records the regulatory team needs from the software.
Pricing depends on scope and is confirmed in a written proposal. Published prices are subject to change.
How Progression’s developers build diagnostic AI
Progression’s own developers build the software around a diagnostic model: data extraction and de-identification, labeling tools, training and versioning, evaluation, integration and monitoring. The team works from New York and meets clinical and regulatory leads on site when asked.
- Write the intended use sentence with the clinical and regulatory leads.
- Map the archives, agreements and de-identification method.
- Build the labeling guide and reader tools.
- Train, version and lock the test sets.
- Validate by site, scanner and subgroup.
- Integrate, run silently, then go live with monitoring.
Who files with FDA?
The manufacturer. Progression’s developers build the software and keep the design, data and test records its regulatory team files from. For strategy before a build, see AI consulting services.
Before the first line of code
Eight answers decide most of what follows. Get them in writing before any model work starts.
- The intended use sentence, signed by the clinical lead.
- The first user of the output.
- Data sources, access owners and agreements.
- The de-identification method for each source.
- The reference standard and who adjudicates.
- The sites held out for testing.
- The regulatory team and the likely submission.
- The updates you expect in the first two years.
What this page leaves out: the Progression Method, proprietary and reserved for clients. Reach out to learn more
Getting found in search
AI, AEO and what is changing
Paid media and lead generation
Websites and design
Choosing and working with an agency
Software and app development
Website design by industry and type
Web development, platforms and hosting
Social, content and brand
By industry and by situation
Frequently asked questions
What is AI diagnostics software?
Who makes the call when AI medical diagnosis software is in the loop?
How many AI-enabled devices are on FDA’s list?
Which specialty leads AI in diagnostics?
What is an intended use statement for a model?
Which FDA submission fits a new diagnostic algorithm?
What is a predetermined change control plan?
Can a cleared diagnostic model be retrained?
How much data does a diagnostic model need?
Can patient images train a model under HIPAA?
What are the two HIPAA de-identification methods?
Where do identifiers hide in imaging files?
What is a reference standard?
Which validation numbers should a buyer ask for?
Does the decision support nondiscrimination rule apply to diagnostic AI?
Can generative AI draft radiology reports?
Can a general chatbot diagnose patients?
What is Good Machine Learning Practice?
How should a hospital test an AI tool before go live?
What does AI diagnostics software cost to build?
Does Progression build AI diagnostics software?
Have a model idea and a data archive?Tell us the intended use sentence you have in mind and where the data sits. You get a free proposal for the build, the records and the price.
Get a free proposalTell us about a project you’d like help with.
Get a free marketing proposal
Tell us what you are trying to grow and we will come back with a plan, not a pitch deck. Same-day reply on weekdays.
