Clinical AI decision documentation without ingesting diagnosis codes
If you are documenting a clinical decision support model, the record Auditrek is built to keep is structural: which model version produced which output, and whether a clinician accepted or overrode it.
Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.
What you are actually being asked for
You run compliance for a hospital group, a diagnostics vendor or a clinical software company, and there is a clinical decision support model in your AI use-case registry that your own gap assessment puts in the high-risk column. Whether the question comes from your board, a customer's procurement team or a supervisory authority, it arrives in the same shape: for this encounter, on this date, what did the model output, how confident was it, and did the clinician accept it or override it.
Right now that answer is assembled by hand. The model output sits in application logs that rotate. The clinician's action sits in the clinical record, behind an access review. The model version sits in a registry your data science team maintains, and the mapping between the two is somebody's recollection of a deployment ticket. A single case can usually be reconstructed if you ask enough people. Reconstructing every case across a review period, against a deadline, is a different exercise, and a reconstruction assembled after the fact cannot itself show that it was not edited after the fact.
The obvious fix, piping clinical events into a governance tool, is the one most compliance officers in this seat refuse, and for reasons they can articulate. Payloads in a clinical decision support pipeline carry health data. Sending them to a third party in order to close a record-keeping and logging gap is a trade people weigh carefully: another processor to assess, another lawful-basis question to put to your own counsel, another store where special-category data could sit longer than anyone intended. Whether those consequences follow in your jurisdiction and under your arrangements is a question for your legal advisers, not for a vendor page. The practical effect is the same either way, which is that the documentation gap stays open because the remedy looks heavier than the problem.
The record you need does not contain the diagnosis
Auditrek is built on one design premise: the record it keeps is structural rather than clinical. It shows that a specific model version produced a specific output at a specific position in a sequence, and whether a human accepted it or overrode it. None of that requires the diagnosis code, the note text or the patient's name. What your own obligations require of you is a judgement for you and your advisers; what Auditrek does is keep the structural record and let you take it to that conversation.
So the ingestion API takes a hash. You compute a SHA-256 digest of the input your model received and send the digest, the model id and version, the output, the confidence, the timestamp and a human override flag. Auditrek stores that as an append-only decision record, hash-chained per ai_system: each record's hash is computed over the previous record's hash plus the payload, so altering one record breaks the recomputation of every record after it.
Where the gate is enforced
The data-minimisation gate is not a clause in a contract, it is server-side validation. POST /api/decisions inspects every payload for recognisable raw sensitive-field patterns before any database write, including ICD-10 style diagnosis code patterns, and rejects the request with a 422 and an error body that names what it found. Subject identifiers are pseudonymised by default.
That matters most during integration, which is when the mistake actually happens. If an engineer wires a coded label into the output field during a sprint, the ledger refuses the write instead of quietly absorbing it, and you learn about it from a failing request rather than from a subject access request eighteen months later.
The workflow Auditrek is built to support
- Record the data processing agreement first
An ai_system cannot be connected until a dpa_confirmation with a signed_at timestamp exists. No API key and no webhook secret are issued before that, and the check runs server-side and returns a 403 if it is missing, so the order cannot be reversed by a keen engineer.
- Register each clinical model as its own ai_system
The early-warning model and the worklist prioritiser get separate chains, because the sequence is what carries the evidence. One system, one append-only ledger, one continuous hash chain to recompute later.
- Map the payload rather than migrating the model
Ingestion is a generic REST and webhook layer with documented payload mapping guides for AWS SageMaker and Azure ML. There are no bespoke platform SDKs in this release, and nothing about how your model serves a clinician changes: you emit an event after the inference has already returned.
- Hash the input on your side of the boundary
Your pipeline computes the SHA-256 digest of the feature vector or prompt and sends the digest. The record can later be tied back to an input you still hold, without the input itself ever crossing into the audit layer.
- Send the override flag with the decision, not afterwards
The human override flag is a field on the same decision record as the output and the confidence value. Acceptance and override are recorded in the same structure, at the same sequence position, rather than joined together later from two systems with different clocks.
- Verify the chain before you cite it
The verification endpoint walks the chain for that record's ai_system from sequence one, recomputes each hash from the stored previous hash and payload, and returns a tamper status of valid or broken with the first broken sequence number. It is a recomputation, not a stored flag you are asked to trust.
- Generate the document for the range you were asked about
The report generator produces PDF and HTML output for a chosen system and date range, mapped to the EU AI Act Article 12 and 13 fields, with one regulation and one format set at launch. Every report carries a visible footer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel.
Why this shape works for a clinical deployment
The audit layer holds no health data
What it stores is the input hash, the model id and version, the output, the confidence value, the timestamp, the human override flag and a pseudonymised subject reference. Raw model inputs never arrive, and subject identifiers are pseudonymised before storage.
The wrong field is refused, not absorbed
Diagnosis code patterns and other recognisable raw sensitive fields are rejected with a 422 before any write. The control is enforced at the ingestion endpoint rather than described in an appendix nobody reads.
The override flag travels with the decision
Each decision record carries the human override flag as a field alongside the output and confidence it relates to, and each record in the chosen date range appears that way in the generated report.
Nothing sits in front of the clinician
Events arrive after your model has returned, over REST or webhook. Auditrek is a record of what your pipeline did, and it makes no clinical determination of any kind.
A dashboard your team can read unaided
Decisions logged, systems connected, last report generated and tamper-verification status, written for a compliance audience rather than for the person who built the pipeline.
One record shape across every system you register
The decision-record structure is the same for any ai_system you register, whatever the model does, with a separate append-only chain per system. A group running several regulated models can read one evidence format rather than one per team.
What a clinical decision record carries
| Field | What you send | Why it is shaped that way |
|---|---|---|
| input_hash | SHA-256 digest of the input your model received | Ties the record to an input you still hold, without the note text or the coded fields crossing the boundary |
| model id and version | The identifier of the exact build that ran | Separates a January model from a June retrain when a case is reopened a year later |
| output | The value your model returned, free of recognisable raw sensitive fields | This is the field the ingestion gate scans, so a coded label wired in by mistake is rejected rather than stored |
| confidence | The confidence value attached to that output | Records how the model qualified its own answer at the moment it answered |
| timestamp | When the inference happened in your pipeline | Places the event on your clock rather than on the ingestion clock |
| human_override | Whether a human accepted or overrode the output | Keeps acceptance and override in the same structure as the output, at the same sequence position |
| subject reference | Your reference for the case | Pseudonymised by default at ingestion, so the chain can be read per case without carrying an identifier |
| sequence_number and prev_hash | Nothing, Auditrek computes both | Each hash is taken over the previous hash plus the payload, so a later edit breaks the recomputation from that point on |
Where the boundaries are
Auditrek documents what your pipeline did. It does not assess whether what you did satisfies any obligation, and the reports it generates are descriptive and structural by design, with the non-legal-advice disclaimer embedded in every one of them.
The practical facts you will be asked about in a vendor review are these. Hosting is Amazon Web Services, in a single region, with no multi-region replication and no choice of data residency. Sign-in is email and password, with auth-gated access for the members of your organisation, and an owner can invite colleagues and revoke outstanding invites; there is no single sign-on. Card payments, subscriptions and VAT at checkout are handled by Stripe, card details stay with Stripe, and billing is tied to the number of connected AI systems. Decision records are append-only and are kept for the life of the organisation that owns them, and you can delete your account and download a JSON export of your data from the settings page.
Things that are not in place, so you can put them in your assessment rather than discover them later: there is no independent security audit or certification such as SOC 2 or ISO 27001, no independent cryptographic review of the hash chain implementation, and no standard data processing agreement offered by Auditrek for its own customers to sign.
If your question is wider than one clinical model, the industry view of the same ground sits on the healthcare documentation page, and the readiness checklist walks the high-risk obligations your gap assessment is probably built from. For a non-clinical example of the same record shape, the algorithmic hiring page shows how the pseudonymised subject reference works per candidate decision.
What logging a clinical model costs
Starter
€1,500 / month
For teams putting their first one or two AI systems under tamper-evident record keeping.
- Connected AI systems: 2
- Decision records: 50,000 / month
- Hash-chained decision records with server-side tamper verification
- Audit reports mapped to EU AI Act articles 12 and 13, in PDF and HTML
- REST and webhook ingestion with mapping guides for SageMaker and Azure ML
- 30-day free trial, no card to start
Growth
€3,500 / month
For teams running several production AI systems at higher decision volume.
- Connected AI systems: 5
- Decision records: 500,000 / month
- Everything in Starter
- Up to 5 connected AI systems
- 500,000 decision records a month
- 30-day free trial, no card to start
Questions
Does Auditrek need access to the clinical record?
No. Your pipeline computes a SHA-256 digest of the input your model received and sends the digest, the model id and version, the output, the confidence, the timestamp and a human override flag. Raw model inputs never arrive, and subject identifiers are pseudonymised by default at ingestion.
What happens if a payload contains a diagnosis code?
The ingestion endpoint inspects every payload for recognisable raw sensitive-field patterns, including ICD-10 style diagnosis code patterns, and rejects the request with a 422 and an error body naming what it found, before any database write.
Can we connect a system before the data processing agreement is recorded?
No. An ai_system cannot be connected until a dpa_confirmation with a signed_at timestamp is recorded. No API key and no webhook secret are issued before that, and the check is enforced server-side.
Does this change the inference path the clinician sees?
No. Events are sent over REST or webhook after your model has already returned its output. Auditrek keeps the record of what your pipeline did and makes no clinical determination of any kind.
Which serving platforms are supported?
Ingestion is a generic REST and webhook layer, with documented payload mapping guides for AWS SageMaker and Azure ML. There are no bespoke platform SDKs in this release.
How do we know a record was not edited later?
The verification endpoint walks the chain for that record's ai_system from sequence one, recomputes each hash from the stored previous hash and payload, and returns a tamper status of valid or broken along with the first broken sequence number. It recomputes rather than reading a stored flag.
What does it cost and is there a trial?
Starter is €1,500 a month for two connected AI systems and 50,000 decision records a month. Growth is €3,500 a month for up to five systems and 500,000 records a month. Both carry a 30-day free trial with no card to start, and billing is tied to the number of connected AI systems.
Where is the data stored?
On Amazon Web Services, in a single region. There is no multi-region replication and no choice of data residency. Card details stay with Stripe and are never stored by Auditrek.