Athletic Heart
CARDIAC SEAL
Adjunctive echocardiography measurement AI (research use)
What it is
Athletic Heart performs automated quantification of echocardiographic studies: ejection fraction, chamber volumes and dimensions, wall thickness, and global longitudinal strain, each computed from the pixel data of the study itself rather than templated. Every study traverses seven ordered safety gates - intake validation, view quality, cardiac-cycle sanity, distribution, physiologic plausibility, conformal confidence, and cross-check - and each must return a pass. The first quarantine halts the pipeline; there is no path from a failed gate to a released result. Every released measurement carries a distribution-free confidence interval, and any study the system cannot pass with sufficient confidence is routed to human review rather than answered. Every decision is recorded in a tamper-evident hash chain, where any later edit, deletion, or reordering is detectable. The regulatory limitations disclosure is a code-level invariant: no licence tier, configuration, or flag can suppress it, and a standing test fails the build if anything tries. The software is not FDA-cleared, it is intended for research use, and a named clinician's signature remains legally required to finish a study.
What is built, and what is ahead
Proven at the bytes: the seven-gate ladder and its fail-closed behaviour, real pixel measurement, distribution-free confidence intervals, the out-of-distribution firewall, the de-identification boundary, the tamper-evident ledger, and a full seeded-defect catch. Explicitly not claimed: clinical accuracy, FDA status, any clinical trial, or a published signed release. Every accuracy figure this system has produced was measured on synthetic phantoms, and the product's own validation protocol says so.
The numbers behind it
What it does
Seven Ordered Gates, Fail-Closed
Intake, view quality, cardiac-cycle sanity, distribution, physiology, confidence, and cross-check must each return a pass. The first quarantine stops the pipeline; there is no path from a failed gate to a released result.
Bounded Autonomy
The system reads every study without a human in the loop, releases the confident ones, and quarantines the rest - and still cannot finish a study without a named clinician's own attestation. Autonomy ends where the legal signature begins.
The Disclosure Cannot Be Turned Off
The regulatory limitations notice is a code-level invariant. No licence tier, environment setting, or account level hides, softens, or removes it, and a standing test fails the build if anything tries. Safety is not a paid feature.
Pre-Declared, Pinned Validation
Before any result existed, the validation targets were declared, cryptographically pinned, and made impossible to move afterwards. A shortfall is recorded as a finding - the bar does not move to meet the result.
Local-First by Default
Out of the box, no patient study and no generated report leaves the machine. Cloud assistance is an operator opt-in, and even then only de-identified numbers cross the boundary.
Fabrication Fails Closed
If the narrative engine invents a number or lets an identifier slip through, the output is discarded and the system falls back to a deterministic report. It is never displayed.
Adversarial Re-Verification
A separate adversarial harness plants known defects into the live source: every one must be caught, and the whole run fails if it misses once. Every gate has been individually forced to fire on a planted fault at its own gate, and every mutated file is restored to the byte or the run stops itself.
Verification Without Trusting Us
Every download can be verified two independent ways using nothing but a standard library - and an unverifiable file is refused outright rather than accepted with a warning. The installed application re-checks its own files at every start-up.
Who it serves
The decision-maker does not change: a cardiologist still reads the study and still signs.
What changes is the provenance of the measurements: every value is computed identically on every run, and a permanent, tamper-evident record preserves exactly what the software did, and when.
It accelerates manual echocardiographic quantification.
Ten parameters arrive already measured from the pixels - volumes, ejection fraction, dimensions, strain, and Doppler values where a spectral strip is present - each carrying its own distribution-free confidence interval, rather than being traced and typed by hand.
Submission evidence is a by-product of operating the pipeline, not a document assembled afterwards.
Traceability, change history, the limitations register, and the decision ledger are produced by operation itself, and every regulatory document on disk is stamped draft-for-counsel rather than presented as settled.
What it decides, and what you decide
Each gate below is a control enforced in the software itself — not a policy statement about how it should be used.
- Ingest, measure, gate, and release-or-quarantine, end to end, with no human in the loop
- The gate ladder's verdicts
- Narrative drafting, reconciled against a deterministic floor before display
- The clinician's signatureReading cardiologist
A study cannot be completed without a named clinician's own attestation, and a signature whose attestation the signer never authored is refused. This is the final safety backstop, and it is legally required.
- Quarantine reviewReading clinician
Any study the gate ladder cannot pass with confidence is routed to a reading clinician rather than answered.
- Regulatory determinationQualified counsel
No clinical claim may be made publicly about this software until a regulatory consultant has ruled in writing. That rule is recorded in the product's own documentation, and this page is written to it.
- Release signing key custodyOperator, offline
The publishing key is held off the build machine. Until it is issued, no signed public release exists - and the release feed states so rather than shipping unsigned.
- Validation data agreementOperator / IRB
Access to real patient data under the declared validation protocol is an operator and institutional-review act; the software never assumes it.
How it is built
- Real pixel-measurement engine
- Conformal prediction + out-of-distribution firewall
- Seven-gate safety architecture
- Hash-chained evidence ledger
- Non-disableable limitations register
- Regulatory-evidence architecture (IEC 62304 traceability)
- Local-first: no study leaves the machine by default
- Fail-closed measurement emission
- Seven-gate safety architecture
- Hash-chained, tamper-evident evidence ledger
- Adversarial harness with a 1:1 seeded-defect bar
- Start-up integrity check scoped to the regulated act
What is running inside it
Every row below is attributed to this system alone and carries the state it is actually in — installed, present in source, or an honest seam.
- Installed
- present and in use in this system's own tree.
- In source
- present in source; not installed, or behind a try/except. Not a shipped capability.
- Seam
- an honest, fail-closed seam. No implementation is wired behind it.
Split-conformal prediction on every emitted measurement
In sourceEach released value carries a distribution-free interval computed by the same pass that produced the value, and the width of that interval is itself an input to whether the study escalates to human review.
Seven-gate fail-closed pipeline
In sourceIntake, view quality, cardiac-cycle sanity, distribution, physiology, confidence and cross-check, in that order. The first quarantine stops the pipeline; there is no path from a failed gate to a released result.
Out-of-distribution firewall
In sourceA study unlike anything the engine was characterised on is routed to review rather than measured with confidence the engine has no basis for.
Keyed hash-chained provenance ledger
In sourceEvery decision is sealed into a keyed, append-only chain whose verifier detects an edit, a deletion, and a reordering as three distinguishable failures. Tamper-evident: the product's own documentation refuses the stronger word for this, and so does this page.
Argon2id credential hashing
In sourceAccount credentials are protected with a memory-hard function chosen over the faster conventional options, pinned to an exact version with the reason written next to the pin.
Ed25519 licensing trust root
In sourceLicence material is verified against a declared signature scheme whose library is pinned directly rather than arriving through some other package's dependency - a gap that was found and closed rather than left to a clean install to discover.
First-party detached release attestation
SeamA download can be verified against a detached signature with nothing but a standard library and no third-party service. The publishing key has not been issued, and until it is, every verifier refuses rather than warns - which is why this is a seam and is named as one.
Classical computer-vision segmentation
In sourceChamber boundaries are found by deterministic image processing rather than a trained on-device network. The product's own honest register states this rather than letting the word AI cover it, and so does this page.
De-identification boundary with an adversarial pin
In sourceOnly de-identified numbers and coded context may cross the boundary to any external assistance, and a standing adversarial test holds that boundary shut rather than assuming it.
Torn-signature permanent marking
In sourceIf an attestation is broken after the fact, the system's own health surface carries that mark permanently instead of returning to a clean state on the next start.
Break-glass privilege with a hard cap
In sourceEmergency elevation exists, is bounded by a ceiling that the emergency itself cannot raise, and is recorded.
How to reach it
Athletic Heart
Built and internally proven; not yet released and not FDA-cleared. Join the waiting list to be contacted when a validation partner programme opens.
Visit echo.aisystemsbymarjerla.com