Works with your existing systems.
DeepMed integrates with Epic, Cerner and NPHIES.
Explore integrationsBring coding, auditing, documentation follow-up and DRG grouping into one workflow. Find the source behind a suggestion, prioritise what needs attention and keep a record of every decision. Agentic AI prepares the work for your team to review.
DeepMed integrates with Epic, Cerner and NPHIES.
Explore integrations
Six steps from the doctor's words to a coded, audited and grouped chart, with people reviewing the results along the way.
Dictate or record the consultation with consent; the doctor approves the draft.
Exports PDF, JSON or FHIR R4Scans become text, and identifiers are masked before any AI sees them.
Suggested codes arrive with the passage behind them, for the coder to accept.
A second AI pass scores the coding and proposes changes for the auditor.
Documentation gaps become neutral, non-leading physician queries.
A rules-based grouper assigns the DRG, final when the chart is locked.
A coordinated workflow combines document processing, de-identification, language models and rule-based validation. Specialist coding, auditing, CDI and DRG tasks pass results between stages, with recorded progress and human review. The stages below show where each task helps your team.
Add one chart, several files at once, or a folder where every file becomes its own chart. Documents are stored on your own SFTP server - the platform keeps the record, your server keeps the file.
Notes written with DeepMed Dictate come in the same way: an approved document exports as a PDF, ready to be coded, audited and queried like any other.
A real text layer is used when there is one, and ignored when it is garbage. Scanned pages are turned upright by an orientation model, checked by actually reading a strip, deskewed, then read by two engines with the better result kept page by page.
Every document keeps its own report: text layer or OCR, which engine won, how many real words came out, rotation and skew, tables, metadata and warnings - so a bad page is visible instead of silently wrong.
Nineteen kinds of identifier - names, addresses, phone numbers, e-mail addresses, health card and MRN numbers, dates, SSN and SIN, card, account and licence numbers, URLs, IP addresses, ages, organisations and locations - plus any pattern of your own. You choose which are masked and what the mask looks like.
Everything downstream, including every AI call, works on the masked text. The de-identification service holds no database connection of its own.
A full clinical NLP pipeline over the de-identified text: sentence and token analysis, dates, ranges and measurements, dictionary lookup against the standard terminology, negation and assertion rules, lab values. Terms are colour-coded in the document, filterable, with the detail behind each one.
It finds 98.7% of the mentions the reference cTAKES pipeline finds, at 98.4% precision - with no Java install to maintain.
Once every document of a chart is de-identified, the whole chart goes to the model in one call. What comes back is checked against your own code list in the chart's scheme - a code that does not exist is recorded, never shown - and each suggestion carries a confidence, a reason and the passage in the note it came from.
Auto Coding accepts them all; Suggested Coding leaves the choice to the coder, who marks primary and secondary and can re-run the AI at any point.
Every coded chart is audited: a summary, a score out of 100, and proposals to add, remove or change a code - each with its reason. The auditor accepts a proposal (and it is applied) or rejects it with a reason. Auto Auditing applies every valid proposal at once.
Charts can also be set to no AI audit at all - the coder still signs off, it is only the machine that stays out of it.
The same call that codes the chart also reviews the documentation against the clinical evidence: what is well supported, what needs clarification, what conflicts, and what looks like it is missing - each with verbatim excerpts located in the document.
The ones worth a specialist's time become prioritised CDI opportunities, with the clinical indicators, what is documented now and the gap between them.
A neutral, non-leading clarification query is drafted from the opportunity and its evidence - never from the whole chart - then edited, saved and sent. The physician opens a secure link, reads the query and its evidence, answers once, and the answer is fed back into the suggested codes under deterministic safety rules.
When the chart is marked as coded, a rules-based grouper - not a language model - assigns the DRG from the selected codes and the encounter details, which AI pre-fills from the documents for the coder to confirm. The DRG specialist accepts or rejects the result, any accepted change regroups the chart, and the DRG becomes final when the chart is locked for billing.
A coder should not have to trust a machine. Every suggestion on the coding screen names the passage in the note behind it; clicking it highlights that passage in the document beside the codes. A suggestion whose code is not in your scheme never reaches the screen at all.
The gap in the documentation is found, turned into a question a physician can answer in a minute, and the answer changes the coding - with a coder checking every change.
Every finding lands in one of four categories - well supported, needs clarification, conflicting, possibly missing - with the excerpts that put it there.
The findings that deserve a specialist's attention, ranked high, medium or low, with the clinical indicators and the documentation gap set out.
A draft written from the opportunity and its excerpts alone - neutral and non-leading by construction - which the specialist edits, regenerates or saves.
A secure link with no login and an expiry. The physician sees exactly what was sent, picks an option or writes a comment, and answers once.
Reconciliation asks the model once - with the physician's response, the query, the opportunity and only the suggestions that bear on it - whether the answer changes the coding. Then deterministic rules take over, not the model:
Coders, auditors, CDI specialists, DRG specialists, supervisors and administrators each get the charts that are theirs - and only the buttons that are theirs.
Whoever uploads a chart is not necessarily whoever codes it. The coder is assigned from the group, sees it in their work list, and is the "who coded it" in every audit and report.
Paged, filterable lists per role, bulk reassignment, and a notification when a chart becomes yours. Supervisors see their group; administrators see everything.
A dashboard over the charts moving through, a package catalogue with subscriptions and reports used, and self-service sign-up with card payment.
Not a policy document - these are properties of how the system is put together.
Every app - this website included - goes through one API with one endpoint, which runs named procedures as a login that can read no table directly.
Every call is checked against the caller's roles before the procedure runs, and a signed-in user can only ever act as themselves.
De-identification happens before coding, auditing, CDI and query generation. A query is drafted from excerpts, never from a whole document.
Reading patient data is logged with the person behind it. The API log keeps ids and statuses; every other parameter value is masked.
Passwords and sensitive settings are stored encrypted with a key that is not in the database, the code or source control.
A query link carries 32 random bytes; only its hash is stored. It expires, can be revoked, and can only ever act on the one query it belongs to.
The platform runs on your infrastructure, with the AI model of your choosing. The heavy work - OCR, de-identification and clinical NLP - happens on your own machines; only masked text ever leaves them, and only if you want a hosted model.
Windows services and local Python services.
OCR uses it when it is there.
Everything goes through the same API.
Masking, prompts, models, intervals.
A demo account opens the whole platform - upload a chart and watch the text come out, the identifiers disappear, the codes arrive with their evidence, and the CDI queries write themselves.
Start a Free Demo Talk to us first