why not AI
We do not compete on the accuracy of the output — we answer for its provability and its attainability
Interpretation is run not by an AI model but by a deterministic pipeline following the rules of the graph. There are no trainable models in the decision path.
Comparison
How we are different from others
- From neural-network interpreters of the electrocardiogram. They explain the output after the fact — with an attention map or a similar case. In a dispute you can produce a rule, not weights.
- From the built-in automatic conclusions of electrocardiographs. They end with a line on the strip: no signature, no document, no history of versions.
- From consumer watches and apps. They give the user a report and stop there; the physician is left with nothing that can be filed into the record.
- From medical information systems where the rules are hard-wired into the code. There, changing a rule is a programmer’s work and a new release of the system; here it is an act over the graph, with a version.
- From remote reading services staffed by other physicians. There the conclusion is given by the service’s physician, not by the one the patient came to. Here the pipeline prepares the draft and your own physician signs it — responsibility and authorship stay with them.
- From cloud services that analyze long recordings and issue a report “on an advisory basis”. Their output is a report the physician confirms. Here the output is a document the physician signs and which goes to the registry.
- From algorithm marketplaces attached to fleets of devices. A marketplace sells algorithms on top of the collection of the signal. Here what is sold is the layer of document, dictionary and evidence, not tied to a device.
- From turnkey complexes with a cloud of their own. A complex sells a device and an archive to go with it. Here it is the path from the recording of any device to a signed document; that is why the makers of such complexes are partners for us, not rivals.
Axes
Eight axes on which we can be compared
| Axis | With us | With others |
|---|---|---|
| A document, not a finding | an instrumental examination protocol that goes to the registry | a confirmation inside their own system, or a report |
| The physician signs | your physician, with their own qualified signature; the key is theirs | another physician at a service, or the confirmation of a report |
| A versioned rule, not weights | behind a finding — the rule, the signal window, the term, the quality of the recording | an explanation after the fact, or criteria with no version |
| Any length in one pipeline | from 30 seconds on a watch to a 24-hour Holter — together with the document | some services claim a wider range of lengths, but the output is a report |
| The coupling of a recording at home and the physician’s workplace | the recording reaches the physician as an analyzed case | data into the record with no physician layer, or an overview in an app |
| A physician’s disagreement repairs the rule | the remark corrects the rule by an act with a version | the retraining of a model, or an edit to the report |
| Regulatory position | a tool for producing documents: the product is not a medical device | many others do hold clearance — an axis where they are stronger |
| The perimeter of the data | processing in Russia, with no training of third-party models | placement and secondary use of the data vary |
Position
Why the decision core is rules and the physician
The rules are defined in the graph in a rule-description language and compiled into executable code: a short response time and low operating costs, with no infrastructure for serving models.
Analysis tasks of different difficulty are solved differently. The assessment of recording quality and artifacts — interference, a detached electrode, baseline drift, insufficient duration — is solved by rules, and the result may be an explicit refusal to process. Morphological analysis — the boundaries of the waves, the intervals, the axis, the rhythm — is carried out by the measurement pipeline. Ruling out plausible mimics — a pacemaker, bundle branch blocks, electrolyte disturbances and drug effects that reproduce the picture of ischemia — is the task where the cost of an error is highest, and the conditions for exclusion are written into the rules explicitly.
The hardest task is solved not by an algorithm but by architecture: not to “guess right”, but to recognize the features reliably, to state the grounds explicitly and to present them to the physician or the patient as an evidence chain.
Position
The place of training data is taken by three mechanisms
- A curated dictionary of the cardiology dialect. The descriptors are laid out across seven layers of analysis: signal, quality, waves, morphology, rhythm, clinical conclusions, patterns; they are bound to the electrocardiographic subsets of international terminologies. Quality is checked term by term rather than declared over a sample.
- The accumulation of knowledge through the authorization of conclusions. The physician’s comment becomes a de-identified remark and lets a specialist correct the rule on stated grounds — with an author and a version.
- The perimeter. The data do not leave the infrastructure of the organization or a cloud on Russian territory, and they do not go into the training of third-party models.
Position
What we deliberately do not measure ourselves by
By the accuracy with which the various aspects of the signal are identified. That contest is run by players with data sets built over many years and with regulatory clearances, and winning it by a frontal attack is not our path.
We do not measure ourselves by the size of data sets: there are no trainable models in the decision path. We do not measure ourselves by effect figures: they will be obtained in the first pilot, by a method approved in advance. We do not measure ourselves by clearances: today the product is a tool for producing documents — the pipeline prepares the draft, and the physician takes the decision.
We answer for something else: the reading of an electrocardiogram is producible, reproducible, and turned into a document a physician has signed. The unique proposition is the combination of a provable output and the attainability of that output on an electrocardiogram of any length.
We answer for the reading: its provability, its reproducibility and its attainability on a recording of any length.