How it works
Ten stages. Quality before measurement, measurement before interpretation, a person before anything is done.
The stages are listed in full below, in order, with the guard each one applies.
Stage by stage
What each stage does, and what it refuses to do.
- Ingestion
A recording becomes a study
Recordings arrive from your machines by REST or CSV upload, with the acquisition parameters they were taken under: sampling rate, lead set, calibration and duration. Nothing is typed in twice.
The guard: A payload outside the configured limits is refused with the reason, rather than accepted and quietly misinterpreted.
- Signal quality
Is this recording usable at all?
Lead integrity, baseline stability, interference, saturation and duration are graded per lead, so an uninterpretable trace is identified while the patient is still on the couch — the cheapest moment to repeat it.
The guard: A recording graded inadequate for automated interpretation produces no measurement and no finding. The system withholds and says so.
- Deterministic analysis
Measurements, from an engine you can name
Intervals, axis, amplitudes, and the rhythm and morphology findings the rule set supports are computed from the stored samples. The arithmetic a reader used to do by hand is done once, identically, every time.
The guard: Every interpretation records the engine version that produced it, so a result can always be traced to the rules that generated it.
- Clinical rules and severity
How much does this matter?
Findings are combined into a severity level against clinical rules, with a confidence figure and an explicit trust floor. Severity is what orders the worklist, so the urgent trace reaches a reader first.
The guard: Below the trust floor the finding is suppressed rather than shown with a low score to be interpreted anyway.
- Explainability
The evidence, stated
Each finding carries the measurement that supports it and the rule that produced it, so a reader can check the reasoning rather than accept a verdict — and a reviewer picking the study up cold can see how it was reached.
The guard: A finding with no stated evidence is a defect, not a display option.
- Narrative reporting
Readable prose, describing only what is established
Where narrative reporting is enabled, a generated report turns validated findings into a paragraph a clinician can read quickly and put in the record. On a deployment without it, the structured findings are the report.
The guard: The report generator is given validated findings as its only input. It is not asked to interpret the signal, and it cannot introduce a measurement that the engine did not produce.
- Clinician review
A person decides
Review, override and sign-off are distinct acts with distinct permissions, so a disagreement is a recorded second opinion rather than an edit. The person who signs is the person the report belongs to.
The guard: A signature is bound by digest to the exact interpretation it covers, and stops applying if that interpretation changes.
- Escalation
Nothing critical waits unnoticed
Studies meeting your configured thresholds become alerts with an owner and an acknowledgement target, which is what turns “somebody should look at this” into a state with a name against it.
The guard: Clinical alert delivery by pager, SMS or email is not implemented: an alert about a finding is raised, owned and escalated inside the system, and the health probe reports that rather than implying a channel exists. The one message this build can send out is operational rather than clinical — notice that an ECG machine is failing, is misconfigured, or has stopped sending — and only where the deployment has a mail relay or an SMS gateway configured. Where it has none, the notice goes to the deployment’s own log, and the record says nobody was told.
- Integration and release
Into the record that already exists
Signed studies are transmitted over FHIR R4 or HL7 v2.5 (MLLP), with an append-only delivery log for every attempt, so the report reaches where the rest of the record lives.
The guard: Only a signed study may be transmitted, and a failed delivery is recorded as a failure rather than retried silently.
- Audit and governance
The account of what happened
Every access and every act is written to a hash-chained, append-only trail, alongside a register of which algorithm versions were validated and for what — the record a quality review or a governance question is answered from.
The guard: The trail is append-only in the database, not by convention, so this claim does not depend on the application behaving well.
Why the numbers can be trusted
A result can always be traced back to the rules that made it.
- The same input produces the same output. Measurements come from deterministic arithmetic over the stored samples, so two runs at the same version agree exactly — and a disagreement between clinicians is a disagreement about interpretation, not about arithmetic.
- A trust floor suppresses rather than hedges. Where a finding’s confidence falls below the declared floor, it is withheld instead of shown with a low score to be mentally discounted. That is the failure mode of every system that shows a number it does not believe.
- Re-analysis is a deliberate act. A newer engine version is re-run over a study with the permission it requires, and the study records which version produced which result — so an upgrade cannot quietly change yesterday’s interpretation.
- Nothing is published without its evidence. Each finding carries the measurement that supports it and the rule that produced it, so a reader checks the reasoning instead of accepting a verdict.
The four commitments
Each stated as a claim, not a value.
The engine decides. The model describes.
No finding in this system originates from a language model.
Measurements, findings and severity are produced by a versioned deterministic rule engine reading the stored samples. A language model is used only to turn validated findings into readable prose — it is never asked what the ECG shows, and a statement it produces that is not traceable to a validated finding is a defect.
Quality is a gate, not a footnote.
An uninterpretable recording produces no interpretation, not a cautious one.
Automated interpretation is published only when the recording passes quality validation. Where it does not, the result is the explicit statement that interpretation was withheld — which is a different thing from a normal ECG, and is presented as such.
A person is in the loop, by permission rather than by habit.
Review, override and sign-off are separately permissioned acts, and disagreement is recorded.
A signature is bound by digest to the exact interpretation it covers and to the person who gave it. An override is stored as an override for ever: the automated finding, the clinician’s finding and the reason, all attributable.
Nothing is asserted without the evidence for it.
Every claim on this site about a capability maps to something the software does.
The limits stated on these pages are not modesty. Automatic renewal, email delivery, invoices and refunds are genuinely absent, and the product reports them as absent in its own health probe rather than leaving them to be discovered.
On screen
What the reader actually opens.
Regulatory position
What this is not
This is a demonstration-grade clinical system. It is not a medical device, it holds no regulatory clearance anywhere, and it has not been validated against a clinical dataset.
It does not diagnose and it does not replace clinical assessment. Every interpretation is decision support for a qualified professional, who reviews, overrides and signs it — and the record shows which professional.
No sensitivity, specificity, agreement or accuracy figure appears on this site, because none has been established for this build. Any such figure would be invented. Read the full position.
See it on your own recordings.
Send a request, and a person reviews it. If it is approved you get a link that creates your own organisation, with the first 2 analysed studies included and no card details — and you are its administrator from the moment you set your password.