The questions a hematologist asks.

Not today, and not for any gene until a prospective trial says it is safe for that gene. Today it predicts the lesions that leave a morphologic fingerprint, states which genes it cannot see, and lets the lab route sequencing accordingly.

You should not, yet. That is why the assay result is bound to the same signed record as the prediction. Over hundreds of patients, on your scanner and your stain, that record becomes the evidence for which genes this gets right in your lab.

It says so, for that gene, on that patient, and tells you to sequence. That is the designed behaviour, not a failure mode.

That is the goal. Axiisium is built peripheral-blood first. Where circulating blasts are present, the initial diagnostic marrow is the procedure it is designed to make unnecessary. Marrow remains standard for aleukemic presentations, MDS versus AML, and remission assessment.

Design partners: labs and pediatric centers with archived diagnostic smears and the molecular results already on file. No new patient, no new needle, and it is how the external validation gets built. kevin.minn@axiisium.com →

For a yes/no diagnosis, often not: morphology plus flow can get you there, roughly 20% blasts that are MPO-positive is AML. But a bare "AML" label is a dead end under modern care. Treatment is genetics-gated, and ELN risk class and targeted therapy require the molecular and cytogenetic context. Axiisium solves for the treatment decision, not just the baseline call.

They are different laboratory workflows on different instruments and timelines. Cytogenetics reads whole chromosomes by karyotype and FISH, in days. Molecular NGS sequences specific genes, in weeks. Fusing them means temporally reconciling two different data structures that arrive at different times, which is exactly the friction the product is built for.

Under WHO 2022 it is diagnostic. "Therapy-related myeloid neoplasm" and "AML with myelodysplasia-related changes" are distinct entities defined in part by clinical history: prior cytotoxic chemotherapy or a prior MDS fundamentally changes the biology and risk of the current leukemia. So it is a genuine feature. Today Axiisium treats clinical history as context the clinician weighs; full clinical fusion is on the roadmap.

It defers, it does not infer. From the fast-available signals, Axiisium produces a provisional, signed call with every assumption flagged in plain language, for example "lineage not confirmed by flow; myeloid assumed", and re-derives it to final when the slow signals arrive, but only while a pending signal could still change the answer. Each state, provisional and final, is its own signed record chained to the last, so how the call evolved is itself auditable. It never fabricates a pending result to force a definitive answer.

No, by design. The morphology model can predict NPM1 status from the smear, but that prediction is used only to rank who to sequence first, never to make the genetics call. The signed WHO 2022 / ICC 2022 diagnosis uses the actual molecular assay result, and a confirmatory assay gates any real decision. The ranking cannot itself drive an unconfirmed clinical action.

Four lesions now have out-of-fold numbers on 129 AML patients: PML::RARA 0.977, CBFB::MYH11 0.922, RUNX1::RUNX1T1 0.882, NPM1 0.819, means over 10 training seeds, patient-grouped nested cross-validation with folds sealed before training. Healthy donors are excluded, because separating a healthy donor from a leukaemic patient is trivial and inflates every figure while looking like lesion detection; with them included these read 0.985, 0.955, 0.933 and 0.907. Earlier single-run figures of 0.930, 0.892 and 0.851 were individual draws rather than means and are superseded, as is an NPM1 figure of 0.894 from a smaller 48-patient run. PML::RARA has since been reproduced on an external cohort at 0.724 on 93 patients neither the model nor its encoder had seen; the other three lesions remain internal only. Every number with its control →

Because the two consensus systems genuinely disagree on the blast threshold when a defining lesion is present. Under WHO 2022, an NPM1 mutation waives the 20% blast requirement entirely. ICC 2022 holds a 10% threshold, and 10 to 19% becomes MDS/AML. Axiisium returns the call under both and shows exactly where they diverge, so the molecular result, not a hard-coded number, decides.

No. Research use only, decision support: a qualified clinician makes the diagnosis. The value is a signed, independently verifiable record of exactly how each call was made, and a confirmatory assay upstream of any real decision, not autonomy.

Every call, and every change to it, is bound to a tamper-evident record: the input digest, the model id and version, the output, the attributed clinician, and a cryptographic signature, each linked to the record before it. The ledger runs in production, and its head is anchored to a public, append-only transparency log with a checkpoint signed by a key that never leaves hardware. You can confirm, independently and with no access to us, that a specific model version on specific inputs produced a given call and that nothing was altered. The production ledger is anchored publicly right now: public log index 2051205018 →. Public anchoring powered by Project AIR.

Yes. A single switch runs the whole system air-gapped: inference, reasoning, and the ledger all run locally with no network calls, so nothing about a case leaves the machine. Every call is still signed and chain-verifiable on the box the moment it is made. The public transparency-log proof, which needs a network the air-gapped box does not have, is deferred rather than dropped: pending chain heads spool locally, export as a portable bundle, and are anchored on a connected machine later. Preview build, research use only. How it runs offline →