Use cases · Tidegate
Use cases

What teams do once they know what their data holds.

Discovery comes first. These are the outcomes the Tidegate engine supports from there, each set out as the problem, how we solve it and why it matters. Each one runs in a cloud account you control, so identifiers and record-level data stay in your environment.

USE CASE 01 OF 05

Safe AI on clinical data

FORAI, data and security teams at health systems, health plans and digital health companies

Request an evaluation
PROMPT · AS WRITTEN3 IDENTIFIERS

Maria LopezNAME seen on 03/14/2026DATE for chest painCONDITION. Lives aloneSOCIAL ISOLATION · ROADMAP on 41 Elm StreetADDRESS; daughter visits weekly.

PROMPT · BEFORE THE MODELSAFE HARBOR MODE

Ana RuizNAME · REPLACED seen in 2026DATE · GENERALIZED for chest painCONDITION. Lives aloneSOCIAL ISOLATION · ROADMAP on [ADDRESS]ADDRESS · MASKED; daughter visits weekly.

Example with made-up data. Safe Harbor mode. Outlined: ROADMAP.
THE PROBLEM

Clinical notes, documents and transcripts are what AI tools need most, and they carry identifiers in free text that field-level filters miss: a name inside a sentence, an address mentioned in passing, a date of admission.

Staff are already using AI tools their organizations have not approved. In a December 2025 survey of 518 hospital and health system workers, 17% said they had used an unauthorized AI tool at work (Shadow AI survey, 2025).

HOW WE SOLVE IT

Find and handle PHI, PII and payment card data before clinical notes, documents or transcripts reach LLMs, copilots or retrieval systems, so teams can adopt AI with sensitive data protected before it reaches the model.

The engine runs in your environment. It labels each identifier it finds by type and location, then replaces, masks or generalizes it under Safe Harbor or Expert Determination mode, before the text moves to a model.

WHY IT MATTERS

Healthcare has had the highest average breach cost of any industry in the annual Cost of a Data Breach Report for 14 consecutive years: $7.42 million in 2025. In the same report, one in five organizations reported a breach involving shadow AI, which added $670,000 to the average cost (Cost of a Data Breach Report, 2025).

USE CASE 02 OF 05

AI-ready data for model development

FORData science and ML teams, and life sciences R&D groups building or evaluating models

Request an evaluation
THE PROBLEM

Models trained on structured fields alone miss much of the clinical picture. FDA has noted that measures of disease severity can be lacking in claims data, and that better ways are needed to examine unstructured data in health records (FDA).

Preparing notes for training means removing identifiers without destroying the clinical content. That work is usually done by hand, one project at a time, and rarely documented in a form a reviewer can check.

HOW WE SOLVE IT

Prepare clinical and claims data for training, fine-tuning and evaluating internal or partner models. Synthetic data for development and testing, released only after memorization tests, with a synthetic data report.

The engine extracts and labels clinical content from notes and documents, with condition status, protects identifiers in your environment and records what it changed, so every training set arrives with its own history.

WHY IT MATTERS

FDA’s 2024 final guidance on EHR and claims data asks sponsors who use AI to curate unstructured data to specify the algorithm’s assumptions, the data used to train it and how it was validated (FDA, 2024). Models meant for regulated use need that record from the first dataset, not after the fact.

CLINIC NOTE · AS WRITTENUNSTRUCTURED

Daniel BrooksNAME, seen 04/02/2026DATE. Heart failureCONDITION, worsening. EFTEST 35%RESULT, down from 45%RESULT. FurosemideMEDICATION increased to 40 mgDOSE after a weight gain of 3 kgSYMPTOM; follow-up in two weeksDURATION.

Clinical facts and timing keptIdentifiers, replaced in the record
TRAINING RECORD · REC_0412READY
CONDITIONHeart failure, worsening LINKED · ROADMAP
MEDICATIONFurosemide 40 mg, dose increased LINKED · ROADMAP
RESULTEjection fraction 35%, from 45% FROM NOTE
IDENTIFIERS2 replaced, each change loggedDaniel Brooks → Owen Hale · Replaced · Expert Determination mode04/02/2026 → 05/09/2026 · shifted +37 days
Example with made-up data. Outlined: ROADMAP.
USE CASE 03 OF 05

Data commercialization and research

FORHealth systems, health plans and data networks licensing or sharing clinical data

Data commercialization and research

License your health data for research and AI, with proof built in.

Prepared under a BAA, protected in your environment, documented for an independent qualified expert and for the recipient’s diligence team. Every release ships with the release pack, including a data card and a pinned manifest.

See data commercialization and research
DATA CARD · EXCERPTDOC 12 · FROM V1.1
SOURCESEHR notes, claims, lab results
TIME SPAN2019 to 2026
POPULATIONAdults seen in outpatient care
TRANSFORMATIONSIdentifiers replaced; dates shifted per person
KNOWN GAPSNo pharmacy claims before 2021
RELEASE MANIFEST · EXCERPTDOC 06
ENGINEengine x.y
RULESrules x.y
MODEExpert Determination
DATERelease 1
Example with made-up data.
USE CASE 04 OF 05

Sensitive data discovery and governance

FORPrivacy, compliance and data governance teams, including state Medicaid agencies and other public payers

Request an evaluation
INVENTORY · 6 SYSTEMS4 FINDINGS
SYSTEMSENSITIVE ATTRIBUTESSTATUS
EHR notesSSNOUTSIDE POLICY
Shared driveDOB · ZIPOUTSIDE POLICY
Faxed lettersNAME · PATIENT · CONDITIONNEEDS REVIEW
Analytics copyMRN · NAME · PATIENTOUTSIDE POLICY
Claims databaseMEMBER ID · DOBIN POLICY
Data warehouseMEMBER ID · PSEUDONYMIZEDIN POLICY
Example with made-up data
THE PROBLEM

Sensitive data does not stay in the systems built to hold it. It ends up in note fields, scanned attachments, exports, shared drives and analytics copies, where tools built for structured data do not look.

Without an inventory of where it lives, risk analyses and access reviews rest on assumptions.

HOW WE SOLVE IT

Inventory and classify sensitive attributes across structured and unstructured data, grounded in Tidegate’s data attribute ontology, for compliance, audit and risk reduction.

The engine scans tables and files alike, labels each sensitive attribute and maps where it sits, so you can apply policy, track change over time and answer auditors with evidence.

WHY IT MATTERS

In January 2025, HHS proposed updating the HIPAA Security Rule to require a written technology asset inventory and a map of how ePHI moves through systems, reviewed at least every 12 months (HHS, 2025). The rule is not final; the federal agenda now targets final action around July 2027 (HIPAA Journal).

The Cost of a Data Breach Report found that breaches involving data spread across multiple environments had the highest average cost, $5.05 million (Cost of a Data Breach Report, 2025).

USE CASE 05 OF 05

Linkage and interoperability readiness

FORTeams joining claims, EHR, lab and note data, or preparing it for exchange with partners

Request an evaluation
CLAIMSJ•••• O•••••NAME••••-••-••DOB021••ZIP
EHRO•••••^J••••NAME••/••/••••DOB021••-••••ZIP
EXPERT DETERMINATION MODE · ONE PERSON · TDG-9K3WLINKED
EVIDENCEName variant, date of birth, ZIP
MATCH0.94
SENT ONOMOP, with labels attached
Example with made-up data. Expert Determination mode: a holder-issued pseudonym. In Safe Harbor mode, linkage codes are random and not derived from identifiers.
THE PROBLEM

The same person looks different in every source: misspelled names, swapped fields, blank values, old addresses. Matches are missed or made wrongly, and few teams measure either.

A RAND study found an average duplicate record rate of 8% in master patient indexes (RAND, via AHIMA). A 2018 industry survey reported 18% (Pew, 2018).

HOW WE SOLVE IT

The next step once data is clean and classified: connect each person’s records within your data, with match quality reported, and export them to OMOP, CSV or Parquet with their labels attached. From v1.1: FHIR with Provenance.

Match quality within your data is reported in every release, so analysts know how far to trust a joined dataset. From v1.1: quality broken out by patient group. On the roadmap: linkage across organizations.

WHY IT MATTERS

FDA’s 2024 guidance names linking patient data across sources as a challenge sponsors must address, and asks that data be accurate, complete and traceable (FDA, 2024). Linkage errors nobody measured carry into every study, model and program decision built on the joined data.

Every use case starts with knowing what your data holds.

We run the Tidegate engine on a sample you choose, scoped in writing, in a cloud account you control, and return release-pack documents 01 to 06 on your own records. Tidegate never buys, sells, holds or brokers data.