Clinical text de-identification API for clinical notes and discharge letters
14-day money-back guarantee · Cancel monthly · Invoice with VAT
Clinical notes, discharge letters, reports and case histories name patients, relatives, clinicians, hospitals, dates and ages. ShinrAI replaces these details with realistic stand-ins before an AI model reads the text. Diagnoses, findings and medication stay readable. ShinrAI 1.6 has been the managed API model since 6 October 2026; its training adds clinical documents.
What is replaced and what stays
| Detail in the letter | Example | ShinrAI |
|---|---|---|
| Patient, relatives and clinicians | “Mrs Emma Clarke” | Replaced, consistently across the letter |
| Hospital, practice, place and address | “St Anne’s Hospital, Leeds” | Replaced |
| Dates and ages | “on 12 March 2026”, “aged 54” | Replaced (ShinrAI 1.3 has no date or age category); optionally keep only the year |
| Identifiers and contact details | Case number, phone number, email address | Replaced when detected as customer or case ID, national identifier or contact detail |
| Diagnosis, findings, medication | “suspected pneumonia” | Kept, so the AI model can work with the clinical content |
The examples are fictional. A rare diagnosis or an unusual course can identify a person even without a name. The linkage-risk option estimates how strongly a text still singles out a person; it is a heuristic, not a count. Assess the residual risk for your purpose.
Pseudonymisation with a mapping you keep
The preset pseudonymize returns realistic stand-ins and, on request, the mapping. Your application stores the mapping apart from the text. Without the mapping, the replaced values cannot be traced back.
- Same person, same stand-in. Within a document, one person keeps one stand-in. A session extends this across many requests, by default for 24 hours.
- Stable later on. Give earlier pairs to a new request, and the same values keep the same stand-ins. The letters of one case stay linkable.
- One-way when you need it. The actions
partialandgeneralizecannot be reversed.partial, for example, keeps only the year of a date.
ShinrAI 1.6 on clinical text, measured and published
ShinrAI 1.6 detects dates and ages, which ShinrAI 1.3 does not. Its training data includes generated clinical documents in six languages and French clinical reports. Innovius measured the release against ShinrAI 1.3 on two public clinical test sets, scoring both models on the same types:
- 250 Spanish clinical cases (MEDDOCAN). On people, places and organisations, ShinrAI 1.6 scores 84.5 out of 100 with exact boundaries and 93.1 when a finding only has to touch the right text with the right type. ShinrAI 1.3 scored 50.7 and 69.8. Of the characters of the personal details, ShinrAI 1.6 masks 95.7 of 100 (ShinrAI 1.3: 55.1).
- 63 German clinic letters (GraSCCo). People, places and organisations: 79.5 out of 100 with exact boundaries, 88.7 with overlap; ShinrAI 1.3 scored 50.4 and 72.6. Of the characters of the personal details, ShinrAI 1.6 masks 94.4 of 100 (ShinrAI 1.3: 52.4).
On the Innovius test with realistic letters, invoices and clinic notes, ShinrAI 1.6 finds about 99 of 100 personal details in English texts, and about 96 of 100 findings are correct; combined, that is 97.4 out of 100 (ShinrAI 1.3: 97.7). The test has 200 texts per locale. On our own held-out medical slices (300 records per language in six languages), ShinrAI 1.6 reads 86.9 out of 100 with exact boundaries, slightly below ShinrAI 1.3 at 87.6.
The score runs from 0 to 100. It drops with every missed detail and with every false finding; specialists call it the F1 score. These measurements describe the model, not a guarantee for your documents. All numbers and methods are on the benchmark board and the languages and data categories page. Pilot on your own letters before you decide.
Use in hospitals, practices and research
- Summarise letters with AI. The pseudonymised letter goes to the AI model. Your application restores the real names locally.
- Research and analysis. Collections of reports and case histories run as Batch jobs at half the price per record.
- Searchable clinical archives. Documents are protected before indexing, so patient names stay out of search indexes and embeddings.
- PDF documents. PDF and DOCX files of up to 10 MB come back as a redacted PDF together with the protected text.
ShinrAI is a technical tool. It makes no clinical decisions and is not legal advice. ShinrAI makes no HIPAA compliance claim. Your legal basis, confidentiality duties and approvals stay with your organisation.
Hosted in Germany or fully on your premises
The managed API runs on STACKIT in Germany and is operated by Innovius UG from Teltow. Synchronous requests are processed in memory, without a content archive and without model training. Asynchronous job content is encrypted and expires 24 hours after submission. For isolated environments, offline packs with signed activation keep all processing on your own hardware. A data processing agreement under GDPR Art. 28 is published. Details are in the on-premises section.
Test with your own letters
Starter costs €39 per month for 50,000 records and includes standard and Batch processing. A record is 1,000 Unicode code points; Batch processing counts half. Larger workloads move to Team, Business or Platform, or to a non-expiring pack, also as an offline pack.
Frequently asked questions
Is this de-identification or pseudonymisation?
With the mapping, it is pseudonymisation in the sense of GDPR Art. 4 No. 5: only the holder of the separately stored mapping can trace the stand-ins back. The actions partial and generalize are one-way. Whether a text is anonymous afterwards depends on the residual risk in your context.
Does ShinrAI keep diagnoses and findings?
Yes. ShinrAI replaces details that point to a person, such as names, places, organisations, dates, ages and identifiers. Diagnoses, findings and medication stay in the text, so a model can work with them. Rare diagnoses can still point to a person.
Is ShinrAI HIPAA compliant?
ShinrAI makes no HIPAA compliance claim. The managed API runs in Germany under the GDPR, with a published data processing agreement. Check your own regulatory obligations before you process patient data.
Can ShinrAI run inside a hospital without an internet connection?
Yes. Offline packs with signed activation run on your own hardware without an outside connection. Processing stays fully local.
Where does ShinrAI process patient data?
The managed API runs on STACKIT in Germany. Synchronous requests are processed in memory without a request-content archive or model training. Job data and API sessions stay only as long as their job or session needs them, by default at most 24 hours. You can also run ShinrAI on your own infrastructure with signed offline activation. The page Data handling and retention lists every retention period.
What does clinical text de-identification cost?
A record is 1,000 Unicode code points; Batch processing counts half. Starter costs €39 per month for 50,000 records. Non-expiring packs start at 5 million records for €1,590, also as offline packs.
Can I get my money back if ShinrAI does not fit?
Not the right fit? Write to support within 14 days of your first plan payment and we refund it in full. Once per company, if you used no more than 10 % of the plan's monthly records.