Death within 30 days
—
Models a clinician can work with.
9risk pathways
11changes you can test
3clinical modules
0black boxes
Most risk scores tell you a patient is in trouble and stop there. Ours tells you which pathway is driving it — volume, access, inflammation, nutrition — and what happens to the numbers if you cap the ultrafiltration rate, plan a fistula, or fix the anaemia. Built for the unit meeting, not for a paper.
Dialysis CoPilot
A patient with a catheter, diabetes and rising inflammation. Tick a change on the left and watch which pathways settle and what it does to the risk of dying or being admitted.
64ydiabetic
catheterHb 9.8, alb 3.2
What you could change
Each one acts on the pathways it actually affects. A change that has nothing to do with the active pathway moves nothing, which is the point.Death within 90 days
—
Death within 1 year
—
Admission within 30 days
—
This is a test patient, not a real record, and the figures are not clinical guidance. What is real is the reasoning: risk is built up from named pathways, and each change acts only where it plausibly acts.
the problem
A risk score that cannot be acted on is just anxiety with a number attached
The sickest get the strongest treatment
A model learns from the record that patients on the aggressive therapy do worse, because that is who got it. Left alone, it will quietly advise withholding treatment from the people it was designed to help.
The outcomes reflect your old practice
Every outcome in the training data happened under the care that was actually given. A model fitted to it predicts what your unit used to do, not what would happen if you changed something.
Changing practice looks like the model breaking
You tighten your access policy, performance drops, and the usual answer is to retrain. That just fits the new habit as tightly as the old one, and nobody learns whether the policy helped.
How ours is built instead
one
Read the patient properly
Labs, session records, medications, access history and events, taken as a trajectory over months rather than a snapshot from the last clinic visit. Missing data is flagged as missing, not filled in silently.
two
Route it through known physiology
Risk is assembled from pathways a nephrologist would recognise and can dispute: volume stress, access failure, inflammation, anaemia and nutrition. Each one has named drivers and thresholds written down.
three
Test the change before you make it
Every suggestion is ranked by how much risk it actually removes for this patient, not by how confident the model feels. When the data cannot support an answer, it says so rather than producing a number anyway.
The conversation shifts from arguing about who is high risk to deciding what to do about it.
module 01
Dialysis CoPilot
working prototype in-centre haemodialysisTwo views. One for the unit, one for the patient in front of you.
unit view
The worklist, sorted by who is actually at risk
Runs the whole cohort across centres and orders it by risk rather than by appointment time. Filters by facility and access type, and shows the mix of what is driving risk across the unit — useful when the answer is a policy rather than a patient.
patient view
Everything for one patient, in one place
Mortality focus with the reasons behind it, next session plan, volume and blood pressure trend, adequacy and delivery, access, labs and medications, events and tasks, and a data quality view that tells you what is missing and what that costs you in confidence.
The nine pathways
Each is built from named clinical drivers with thresholds you can check and argue with.
Volume and fluid removal
Interdialytic weight gain above 2.5 kg and ultrafiltration rates above 10 ml/kg/hr. Residual urine output softens it.
Intradialytic hypotension
Crash frequency, worsened by high ultrafiltration and limited cardiac reserve. The single biggest driver of admissions.
Missed and shortened sessions
Amplifies volume, potassium and adequacy all at once. Usually a transport problem, rarely a motivation problem.
Adequacy and delivery
Kt/V shortfall and the gap between prescribed and delivered time. Silent until it is structural.
Access failure
Catheter and graft penalties, venous pressure trend and alarm frequency. Rising pressures are the early warning of stenosis.
Metabolic instability
Potassium above 5.2, phosphate above 5.5, bicarbonate below 22, with a diabetes increment.
Inflammation and infection
Catheter presence, bloodstream infection events and falling albumin. Carries the heaviest weight into mortality.
Anaemia and nutrition
Haemoglobin below 10.2 and albumin below 3.4. Second heaviest, and among the most fixable.
Cardiovascular reserve
Heart failure, age and current volume stress. Changes how much the patient tolerates rather than acting alone.
When risk is high, it shows the mechanism
Not a bar chart of feature importances. The actual chain, with how strongly each link is firing for this patient.
where this stands
Prototype today. Registry next. Nothing claimed before it is earned.
The modules run and the reasoning holds up on test patients built to stress it. That is the honest state of things, and we would rather say so than imply more.
- NOWWorking prototypes, tested against constructed cases across all three modules.
- NEXTValidation against real departmental records, retrospectively.
- THENConnection to hospital systems so the data arrives without manual entry.
- LATERProspective evaluation across centres, against pre-agreed questions.
Watching for failure without waiting years
Hard outcomes in kidney disease arrive far too late to tell you a model has stopped working. We track the gap between what was predicted and what the record subsequently shows, so a drifting model surfaces early. This is the founder's published research area.
The reasoning is written down
The pathway map is a document that clinicians review and sign off, versioned like a protocol. When you disagree with a number, you can find the assumption behind it and argue with that instead of with a black box.
It stays your decision
Nothing here prescribes. It makes trade-offs visible so the team can weigh them, and it is explicit about what it does not know. Clinical responsibility does not move.
No real patients yet
Every figure on this site comes from constructed test cases. Nothing will be described as an effect on patient care until it has been tested forward, in a defined population, against a question agreed in advance.
If you run a unit, we want to be told we are wrong
What these need now is real data and clinicians willing to dispute the pathway maps. We are looking for dialysis networks, stroke services and interventional radiology departments — and for anyone who would rather have the assumptions on the table than buried in a model.
hello@kuviyam.aimodule 02
Stroke Pathway CoPilot
In acute stroke every decision trades something against something else, and all of them are made in minutes. Transfer or go direct. Imaging certainty or delay. A stent that stays open or a bleed. This lays the trade-off out while there is still time to act on it.
What it covers
working prototype hyperacute strokeDecisions it covers
- routingTransfer to a thrombectomy centre, or go direct
- treatmentThrombolysis before clot retrieval, or retrieval alone
- imagingFull perfusion workup, or straight to the angiography suite
- complexTandem lesions, and whether to stent acutely
- large coreWhether to intervene when much of the territory is already lost
- unknown onsetWake-up stroke, where imaging decides eligibility
What it estimates
- timeHow much delay the pathway choice actually adds
- tissueHow much brain is lost while that happens, given collateral status
- harmRisk of symptomatic bleeding under each strategy
- outcomeChance of walking out independent at three months
Collateral status is treated as what it is in practice: the thing that decides how quickly delay turns into infarct, and therefore how much any of these choices matter for this patient.
Come and stress-test this
The most useful thing a stroke physician can do with this is try to break it. If the reasoning is wrong for your service, we want to hear exactly where.
hello@kuviyam.aimodule 03
Body Intervention CoPilot
Planning an ablation means holding the patient's reserve, the anatomy and the lesion in mind at once, then running the trade-offs in your head under time pressure. This makes them visible before the needle goes in, and lists the mitigations in order of how much risk each one actually removes.
What you can change
What it estimates
before the meeting
Cohort triage
Works through the department list and flags the cases where reserve is marginal or where the standard approach is likely to run into trouble. Sorted by risk, not by referral date.
in the meeting
Trade-offs on screen
Change the energy, the margin, add hydrodissection, correct the coagulation. The numbers move while the team is still talking, so the discussion is about the plan rather than the score.
after
Follow-up that closes the loop
Response assessment and surveillance in the same place as the plan, so what was predicted can be checked against what happened.
It does not tell you what to do. It makes the trade-off visible so you can decide with your eyes open.
Come and stress-test this
The most useful thing an interventional radiologist can do with this is try to break it. If the reasoning is wrong for your practice, we want to hear exactly where.
hello@kuviyam.ai