A patient with suspected AML has a smear under the microscope within hours. The genetics that classify the disease and choose the therapy take days. Treatment for a fit patient cannot always wait, and a targeted drug cannot be given without its mutation confirmed. Every hour in that gap is a decision made on incomplete information. Every marrow drawn to close it is a procedure the patient feels.
For each target lesion, three possible states.
With a calibrated probability.
With a calibrated probability.
The model declines to call it, for that gene, on that patient.
The third state is the product. A model that knows the limit of what morphology shows, per gene, per case, is what lets a lab route sequencing to the patients and genes that need it instead of running everything on everyone. Abstention is implemented as conformal prediction: the abstention threshold is fit on held-out calibration data, so the coverage guarantee is a property of the method rather than a promise.
Not every mutation reshapes a cell, and not everything we intend to build is built. This table is updated when the numbers change and never before. Four lesions now carry out-of-fold numbers on 129 AML patients; the rows below them still do not, and say so.
| Lesion | Morphologic basis | Status today |
|---|---|---|
| PML::RARA (APL) |
Hypergranular promyelocytes, faggot cells | Demonstrated in preliminary testing AUROC 0.977 (0.948–1.000), out-of-fold. External: 0.724 on 93 unseen patients |
| CBFB::MYH11 (inv16) |
Abnormal eosinophils | Demonstrated in preliminary testing AUROC 0.922 (0.859–0.985), out-of-fold, no external cohort |
| RUNX1::RUNX1T1 (t8;21) |
Salmon granules, perinuclear hof | Demonstrated in preliminary testing AUROC 0.882 (0.809–0.954), out-of-fold, no external cohort |
| NPM1 | Cup-shaped nuclei, condensed chromatin, perinuclear clearing | Demonstrated in preliminary testing AUROC 0.819 (0.726–0.912), out-of-fold, no external cohort |
| Per-gene abstention | Conformal prediction over calibrated probabilities | In development Running, but not yet measured against an external cohort |
| FLT3-ITD | Weak independent signal; published AUROC 0.74 to 0.82 elsewhere | Planned No public labelled data exists. Would ship as a fast-track flag, never as a skip |
| TP53, IDH1, IDH2, RUNX1, ASXL1, splicing factors, KMT2A |
No morphologic fingerprint demonstrated by anyone | Cannot see Always sequenced |
PML::RARA is the urgent one, and it is also the strongest. APL is a medical emergency and ATRA is already started on morphologic suspicion in practice. A systematic Day-0 flag for it is the single most time-critical thing this system can do, and at 0.977 out-of-fold it is the lesion the model reads best. It is also the one lesion tested outside its own dataset: 0.724 AUROC on 93 patients from a cohort the model and its encoder had never seen, which establishes the signal survives the transfer without measuring how well, on ten positives.
Peripheral blood is the cleaner substrate for a model, because cells are dispersed instead of crowded, and the sample is a blood draw instead of a procedure. Axiisium is built peripheral-blood first.
For the majority of adult presentations with circulating blasts, the initial diagnostic marrow is the procedure this is built to make unnecessary. In children, it is the repeat marrows that hurt most, and they are the longer road.
Marrow remains necessary when circulating blasts are scarce, when the question is MDS versus AML and cellularity matters, and for remission assessment.
A real workup does not arrive at once. The smear is under the microscope within hours; the genetics take days. Axiisium makes the Day-0 prediction, names the genes it cannot see, and binds every later result to the same record.
NPM1: predicted present, 0.87. PML::RARA: predicted absent, 0.98. FLT3-ITD: cannot see, sequence.
Myeloid lineage confirmed, not ALL. Bound to the same record. NPM1 prediction unchanged.
NPM1 PCR run first, ahead of the full panel. FLT3 assay ordered. Targeted-therapy eligibility flagged for the treating team.
The assay agrees with the Day-0 prediction. Under WHO 2022 the NPM1 mutation waives the 20 percent blast rule; under ICC 2022 the 10 percent cutoff is met at 15 percent. Final: AML with mutated NPM1. The full panel continues for the genes morphology could not see.
This is the intended flow. The rules engine and the signed record shown here run today. The Day-0 genotype call is at the status given in the table above.
Morphology is what Axiisium reads from images. Immunophenotype, cytogenetics, molecular results and clinical data are taken as structured inputs. When an assay returns, it is bound to the same record as the prediction, so the case shows what was predicted, what was confirmed, and where they differed.
Working todayA predicted lesion only matters if it changes the classification. Axiisium applies WHO 2022 and ICC 2022 and shows where a defining mutation overrides the blast threshold under each, and marks a call provisional while a pending result could still change it.
WHO 2022 + ICC 2022A lab that runs NPM1 first, or defers a gene, on a model's word needs to prove what the model saw. Each prediction, state change and confirming assay is bound to a tamper-evident record: exact model version, inputs, attributed clinician, cryptographically signed and linked to the one before it. Change one byte and verification fails.
Signed · chained · anchoredOne switch puts inference, reasoning and the ledger on a fully local path with no network. Every call is still signed and chain-verifiable on the machine that made it. Public anchoring is deferred through a portable bundle, not dropped.
AXIISIUM_OFFLINE=1 · previewThe production ledger head is anchored on a public transparency log. Read the recorded hash and confirm it matches, with no access to Axiisium and no trust in us.
Public anchoring powered by Project AIR.