SPECTRAL · CLINICAL DOCUMENT INTELLIGENCE
Clinical decisions are written in letters. Spectral turns them into data.
Spectral is Toca Health’s document intelligence software for the NHS. It reads discharge summaries, clinic letters and referrals, and extracts the decisions inside them, things like fitness for discharge, follow-up and coding, as structured, auditable data your systems can act on. It runs inside your own network, and no patient data leaves the Trust.
THE PROBLEM · NOT WHERE THE DATA IS
The most valuable information in a Trust is locked in prose.
Fitness for discharge, two-week-wait status, agreed follow-up: these decisions sit in the paragraphs of consultant letters, and no reporting tool or data platform can query a paragraph. Today a person reads each letter and types what it says into a system. That manual step is the bottleneck, and it is why so much of what a Trust knows never shows up in what a Trust can measure.
Spectral works on the other side of that line. It doesn’t move your records or replace what reads them. It turns the correspondence you already hold into structured, scored, auditable decisions, where it sits, so the systems you already own have something to act on.
HOW IT WORKS · ONE LETTER, SEPARATED
White light in. A spectrum of decisions out.
Spectral takes one letter and extracts each decision in it as a separate, structured finding, with the source passage attached. Each finding is assessed independently. Clear-cut findings are actioned; ambiguous ones go to a clinician. The letter is never treated as one all-or-nothing judgement.
Scroll the diagram →
CONSENSUS VALIDATION · THE MECHANISM
A jury of models. One accountable verdict.
Every extraction is put to more than one model. Unanimous findings proceed. If any model dissents, the case stops and goes to a clinician with the disagreement shown. Each verdict is recorded in the audit trail alongside the source text it was drawn from.
Read
The letter is parsed in place, inside the Trust network. Nothing is uploaded and no external service is called.
Separate
Each decision and data point is extracted independently, with the passage it came from attached as evidence.
Put to the jury
More than one model assesses every extraction. Agreement and disagreement are both recorded.
Route
Unanimous findings proceed at the threshold the Trust has set. Anything short of that goes to a person.
Figures from live Spectral processing at Royal Berkshire NHS Foundation Trust to 4 August 2026. The review rate is measured over a 213,000-document sample to 8 July 2026, of which 24,000 were routed for human review. Results vary by document type and by the confidence thresholds each Trust sets.
Spectral deployments
WHAT IT DOES · THREE CLINICAL FUNCTIONS
Built for specific jobs, not general intelligence.
Spectral is deployed as named clinical functions, each with defined inputs, defined outputs and its own evidence base.
Discharge decision analysisLive at Royal Berkshire
Reads consultant correspondence at scale, identifies which documents are discharge summaries, and separates the decisions inside them, fitness for discharge, destination, follow-up, medication, onward referral. Where the jury is unanimous the discharge proceeds; where it isn’t, a clinician reviews.
In live useClinical codingInpatient & outpatient
Reads the clinical record and proposes ICD-10 and OPCS codes with the supporting passage attached, in around 38 seconds per suggestion. Coders review and confirm; nothing is submitted on the model’s word. In build at Royal Berkshire and Morecambe Bay, where hardware is installed and initial testing has completed successfully.
In build, not yet liveReferral cohortingAlongside eTriage
Classifies incoming referrals by specialty, urgency indicators and pathway suitability at the point of intake, and surfaces the relevant history to the clinician answering an Advice & Guidance request, so triage time goes to the cases that need judgement.
Evaluation stageWHERE IT SITS · ACROSS THE PATHWAY
One capability, used at several points on the pathway.
Spectral isn’t a stage in the pathway, it’s a capability that surfaces wherever a decision is locked in text. It reads from and writes back to the systems the Trust already runs.
Patient pathway view, longitudinal record
Primary care
GP contactReferral letter authored and submitted through NHS e-RS.
Referral & triage
GatewayReferrals read, prioritised and routed to the right clinic.
Waiting list
BookingLists managed, clinics filled, RTT clock tracked.
Outpatient
Consult & treatClinical activity captured, coded and reconciled.
Discharge
CompletionDischarge letters analysed, coded and routed.
Follow-on
Care transferOnward referrals managed and handed off cleanly.
Insights & AI
- Referral cohorting
- Discharge decisions
- Dynamic scheduling
- Patient flow
- Predictive analytics
Data logic
- Clinical coding
- PDS / demographics
- RTT pathways
- Data quality
Core systems
- NHS e-RS
- PAS
- EPR
- Order comms
- Integration engine
Highlighted capabilities are Spectral’s, live or in build, as marked in the cards above. Everything else is your existing estate, which Spectral reads from and writes back to.
SOVEREIGNTY · WHERE THE PROCESSING HAPPENS
The data boundary is the Trust boundary.
Spectral runs on your infrastructure. Patient data is processed inside your environment and does not leave it. No identifiable information goes to an external AI service, and models are deployed as fixed, versioned artefacts, never trained on Trust data. That keeps the DPIA straightforward: the data-flow boundary and the organisational boundary are the same line.
On-premiseYour data centre
Runs inside the Trust network on your own hardware. No external calls in normal operation. The deployment model at Royal Berkshire and Morecambe Bay.
Your cloud tenancyNHS-compliant
Deployed into a cloud tenant you control, Azure, HSCN-connected or equivalent. Your tenancy, your keys, your access policy.
HybridSplit by sensitivity
Clinical processing on-premise, with orchestration and reporting in a secure environment where that suits your estate better.
ASSURANCE · EVERY LAYER, NOT ONE BADGE
No single certificate makes clinical AI safe. So we evidence every layer.
Assurance for clinical AI is layered: device regulation, clinical safety standards, data protection, information security and human oversight each have their own requirements and their own evidence. The table below sets out our position on each, and where a question is genuinely open, it is named rather than glossed.
| Layer | Position |
|---|---|
| DTAC | Completed DTAC evidence provided, covering clinical safety, data protection, technical security, interoperability and usability. |
| DCB0129, manufacturer | Safety case maintained by Toca as manufacturer, with a named Clinical Safety Officer. |
| DCB0160, deploying organisation | We support each Trust’s own safety case, hazard workshops and local safety documentation. |
| MHRA device status | Registered as a Class I medical device. Classification for newer use cases is confirmed with the Authority, and with the MHRA where appropriate, before deployment rather than assumed. |
| ISO/IEC 42001 | Certified at company level, external assurance across the full AI lifecycle. |
| Data protection | On-premise processing, no data egress, no retraining on Trust data. We support each Trust’s DPIA. |
| Information security | ISO 27001, DSPT and Cyber Essentials maintained; penetration testing evidence available. |
| Post-market surveillance | Ongoing monitoring of model performance, with incident reporting aligned to LFPSE and MHRA routes. |
| Human oversight | Consensus Validation operationalises NHS England’s principle that a clinician verifies AI output, a person owns every clinically significant decision. |
FAQ · GOVERNANCE, DATA AND PROCUREMENT
What Trusts ask us first.
Does any patient data leave our network?
No. Spectral processes documents inside your environment, on your infrastructure. No identifiable data is sent to an external AI service and no third party sees your records.
Do you train models on our data?
No. Models are deployed as fixed, versioned artefacts. Your documents are used to produce output for you, and for nothing else.
Does anything get actioned without a clinician?
Only where your Trust has set a threshold allowing it, and only where the model jury is unanimous. Where any member dissents, the case stops and a clinician reviews it with the disagreement shown. The thresholds are yours to set and yours to change.
Does this replace our EPR or data platform?
No, and it isn’t an alternative to one. Spectral reads the correspondence you already hold and writes structured output back into the systems you already run, including a federated data platform if you have one. It solves a different problem: turning text into data.
What do we need in place to read our documents?
Less than most Trusts expect. Spectral needs to reach the correspondence and to write results somewhere. Where you have an integration engine and APIs, those are what we use, every time: a supported interface is always the better route. Only where one isn’t available, or the work sits behind a vendor change request, do we automate at the user interface layer instead, under a service account with permissions you set. There is no requirement for a data platform, a warehouse or a new interface before you can start.
How does it relate to eTriage?
Spectral is the document intelligence layer; eTriage is the referral and Advice & Guidance workflow built on top of it. One platform, one governance pack, one support route, see governance for the detail.
How do we buy it?
Through G-Cloud, as a standard Call-Off Contract in most cases. Deployment starts with a scoped evaluation on your own documents, in your own environment, before anything touches a live pathway.
Run it against your own letters.
Every deployment has started the same way: a scoped evaluation on your documents, in your environment, with your clinical leads reviewing the output.