VitalMatch is a project-scoped research platform for organ-donor screening — cohort browsing, configurable DICOM de-identification, a 2D/3D CT viewer, model training, and a Python SDK on top of the UNOS STAR registry.
Service healthy
·
GET /api/v2/health
·
v2.0.0
Everything a research team needs to take raw DICOM and a STAR release through cohort definition, training, and reproducible export.
Each research question gets its own container — subjects, scans, metadata, annotations stay project-scoped.
Donor lookup, recipient outcomes, cohort browse with filter, facet, and CSV/Excel/Parquet export.
Reader / editor / admin roles with an immutable audit log of every action.
HIPAA Safe Harbor, Strict Research, or your own DICOM profile — per project.
2D slice browser with window/level controls and a marching-cubes lung mesh for any scan.
Project-scoped CSV/JSON/printable exports and binary-classifier training jobs.
pip install vitalmatch-sdk, mint a token, reach every endpoint from a notebook.
dd-cfDNA-style post-transplant monitoring scaffolding and recipient graft-health LLM tool.
LLM conversational layer with multi-tool registry access, persisted across sessions.
A running series of aggregate analyses of the UNOS STAR registry that trace one arc: a donor-lung supply that keeps growing yet is increasingly discarded; the blood-type, size, sensitization, and allocation filters that decide who gets an organ; and the graft outcomes and waitlist mortality that make every usable lung count. Together they map where objective, offer-time donor-lung assessment moves the needle. No PHI — each brief is self-contained and shareable.
Among 42,925 adult lung transplants since 2005, the 819 recipients with systemic sclerosis survive as long as everyone else after adjustment (HR 1.01, 0.91-1.12), but their risk is concentrated in the first year (HR 1.21), with more severe PGD (42% vs 32%) and early rejection. Bronchiolitis obliterans is not more common. No regimen can be shown to be better for scleroderma specifically; tacrolimus is associated with better survival than cyclosporine in scleroderma and everyone else, and scleroderma does not modify the induction or antimetabolite associations. Built from 1,825,759 recorded values across four STAR tables.
Open reportLAS (2005) and CAS (2023) were each justified by two promises: fewer waitlist deaths and better post-transplant survival. In matched three-year windows either side of each change, both delivered the first and neither the second — 1-year survival 83.2% to 83.6% at LAS, 89.0% to 90.1% at CAS, while waitlist mortality fell 28% at the CAS transition. Crucially, waitlist mortality had been rising through the LAS era (11.5 to 13.1 per 100 patient-years), so CAS breaks an adverse trend rather than continuing a good one. Transplant volume rose over the same period, so no causal claim is made.
Open reportFresh and salt water injure the lung differently, so partitioning drowning donors by water type is the obvious next question. It cannot be asked: searching all 75 table sheets and 49,143 variable rows of the STAR dictionary returns a kidney pump manufacturer, a country name, and the drowning code itself. Sex has no effect (7.1% vs 7.5% lung recovery, p=0.68); the recovery deficit is significant in all six age bands, and 44% of these donors are aged 0-4. What the registry does record is the airway: drowned donors have the highest abnormal-bronchoscopy rate (34.9% vs 29.1%), driven by purulent secretions rather than documented aspiration.
Open reportGunshot-wound donors give up a lung 2.29 times as often as other mechanisms — 38.6% against 16.9% — and are above par on every other organ too. Age does not explain it: the advantage holds in every band, and age-standardised observed/expected is 1.69. Unlike the asphyxiation utilisation gap, this survives stratification by donation pathway. Early graft function is identical; one-year graft survival is 86.3% against 85.0%, significant but too small to act on. The likely explanation is not the injury but what surrounds it: youth, no chronic disease, isolated trauma and a short ventilated course.
Open reportDonors who died by drowning yield kidneys at 95.5% — the highest rate of any mechanism of death examined — and hearts well above average, but lungs at 7.2% against 18.7%. The deficit survives conditioning on the donation having proceeded and survives age standardisation, leaving 220 donors who gave a lung against 431 expected. Recipient outcomes were reported equivalent in a propensity-matched analysis five years ago. The gap is narrowing and ex-vivo perfusion is already used here more than for any other mechanism — but lungs never recovered have no outcome, so whether this is right selection or recoverable supply cannot be settled from the registry.
Open reportHypoxia and aspiration make asphyxial donors the obvious candidate for a death mechanism that damages lungs. Across 12,509 such donors the premise fails four independent ways: utilisation, severe PGD (33.6% vs 32.7%, p=0.61), one-year graft survival, and all fourteen donor-CT features. The one significant result — a 1.3-point utilisation gap — is Simpson's paradox: it vanishes on stratification by donation pathway and reverses within brain-death donors. The registry also has no strangulation code, so the question that started this cannot be answered directly.
Open reportCAS is a priority ranking, so asking whether the total predicts survival is a category error — it scores 0.455, below 0.5, because sicker candidates rank higher and then do worse. But one component exists to predict post-transplant benefit, and it works: C-index 0.579, with 1-year mortality falling 25.2% to 12.0% across its quartiles. No component predicts PGD. The hard limit is follow-up: CAS began March 2023, so 4,379 transplants have a year observed and none has five — an interim read on an instrument built to predict five-year benefit.
Open reportSTAR types nearly every donor for HLA — 99.4% at HLA-A — but almost never by DNA. Recipient molecular typing rose from 0.0% before 2005 to 15.0% today while the donor side reached only 3.3%, and epitope algorithms need both. PIRCHE-II is eligible for 59 pairs (0.1%), and restricting to recent transplants does not rescue it. Antigen mismatch is computable for 84.3% but does not predict severe PGD — which fits the mechanism, since PGD is ischemia-reperfusion injury, not alloimmunity.
Open reportCause of death for 4,845 deaths in the graded cohort, 98% recorded. Severe PGD adds 30 multi-organ-failure, 19 infection and 18 graft-failure deaths per 1,000 recipients in the first year. Read as shares instead of rates, infection looks less important in severe PGD than without it (20.5% vs 26.3%) while actually killing nearly twice as many — the brief shows both views and why only one answers the question.
Open reportGrading primary graft dysfunction by severity changes what it predicts. Moderate PGD is indistinguishable from none at every horizon (five-year survival 62.7% vs 63.1%); severe PGD costs about eleven points, all lost in year one — an adjusted hazard ratio of 2.51 through the first year, falling to 1.12 and losing significance beyond five. A published analysis of the same registry reports no effect after adjustment; we do not reproduce it, and show why the specification is suspect.
Open reportA European series argues for transplanting the liver first in combined lung-liver cases. U.S. data show that practice emerging — 38 of 196 cases carry prolonged lung ischemia (10.5 h vs 4.8 h, median year 2024), matching the paper's 10.2-vs-5.2 signature. Allocation data narrows it further: in the CAS era the two groups are indistinguishable on every measure of recipient severity, and the one component that separates them is placement efficiency — these organs travelled further. Sequence itself is never recorded, so the proxy tracks distance but cannot see implant order, and the outcome question stays open.
Open reportLung-liver recipients appear to outlive lung-only recipients (62.2% vs 55.9% at five years) — but that advantage is era confounding, not procedure: dual-organ volume is concentrated in recent years, when lung survival improved. Matched to 2015+, the two are level (61.0% vs 59.7%). The real signal is an early penalty — one-year survival 80.7% vs 88.9% — after which trajectories converge. Heart-lung has improved most: 48.0% all-era vs 61.2% since 2015.
Open reportDICOM header audit of all 16,404 CT archives: a CT_ filename is not a chest CT (83% contain a chest series, the rest are head/abdomen/trauma studies) and most reconstructions are too thick for quantitative work — only 29% of series are ≤1.5 mm. Combining both filters leaves ~9,300 analysis-ready archives, about half the headline count. BodyPartExamined is missing in 38% of series, so header-level filtering at ingest is mandatory.
Open reportSurvival is the abundant, long-horizon outcome: 50,247 recipients (99.8%) have follow-up, with assessable survival for 46,254 at 1 year down to 32,237 at 10 years. Crude survival rose 77%->87% at 1 year from the pre-LAS to LAS era; SRTR death linkage recovers 988 deaths OPTN follow-up missed. Severe PGD@72h ~6x the 30-day mortality (7.1% vs 1.1%) with a gap that persists to 5 years — figures since superseded by the graded, day-3-landmarked analysis. The label layer to anchor models.
Open reportSelection audit of the DICOM→STAR crosswalk: donors with a linked chest CT were transplanted 91.5% of the time vs 15.2% without one, so the imaging set is almost all accepted lungs. The discard arm is just 790 donors — older, DCD-heavy, and declined mostly for physiologic 'poor function' — leaving only ~182 image-assessable declines. Scope companion to the Discarded-Lung CT Cohort brief.
Open report1,774 discarded lungs (1,029 donors, ~19% of all discards) carry a donor chest CT — a labeled validation set of declined organs. Discarded lungs had near-normal oxygenation (median P/F 400 vs 444 transplanted), ~1,500 declined for 'poor function' despite acceptable gas exchange: an objectively-questionable discard pool CT can adjudicate.
Open reportSevere primary graft dysfunction at 72 h — VitalMatch's model target — affects ~20% of bilateral lung recipients (ISHLT grade-3 proxy), rising over time and increasing with ischemic time (16→23%), DCD (25% vs 19%), and donor age. The drivers are all offer-time knowable.
Open reportCandid status of the severe-PGD@72h classifier: clinical LR AUROC 0.68 (cross-seed median) on a small held-out set (251 cases, 22 events), beating published donor scores (~chance–0.64). RF looks higher but is unvalidated; imaging adds no value yet. Research use only — preliminary.
Open reportRegistry-scale audit of U.S. donor-lung non-utilization: an 8.1% discard rate rising to 14.4% in 2025, concentrated in DCD and older donors, driven by subjective “poor organ function” calls.
Open reportDecomposes the rising lung-discard rate: DCD donors climbed from 1% to 24% of the lung pool and carry a ~5× discard rate, so much of the headline trend is compositional, not a worsening of any single donor type.
Open reportRepresentativeness check: the ~9,300 donors with linked chest CT match the national lung-donor pool within a few points on sex, DCD, and smoking — establishing external validity for models trained on the cohort.
Open reportMachine-perfusion adoption reached ~5% of lung donors by 2025; perfused donors skew marginal (mostly DCD) yet still convert to transplant — a rescue channel that pairs naturally with objective CT triage.
Open reportLung discard varies 3.3× across the eleven OPTN regions (4.6% to 15.0%) — practice variation, beyond donor case-mix, that a consistent objective read could compress.
Open reportOne-year graft failure is U-shaped in cold-ischemia time — lowest at 5–6 h, rising beyond 10 h and (via confounding) under 4 h. The long tail is the modifiable target for offer-time efficiency.
Open reportLung-graft survival fans apart with time: youngest donors 85%→51% (1→5 yr), older donors trail by ~6 pts at 5 yr, and DCD grafts diverge late (40% vs 49%) — donor quality is a long-horizon cost the 72-hour window misses.
Open reportSince continuous distribution (Mar 2023), lung transplant volume rose ~30%, median cold ischemia climbed 5.8→8.9 h, lungs >8 h ischemia 17→57%, and DCD use 7→19% — broader sharing traded geography for time.
Open reportTwo hard filters gate every lung offer before quality: 91% of transplants are ABO-identical (+9% compatible) and 72% within ±10 cm donor–recipient height — compounding constraints behind 'no recipient located' discards.
Open reportMost lung recipients are unsensitized, but the cPRA 80–100% tail (~1%) waits ~1.8× longer (72→127 days median) — the minority where epitope-level (PIRCHE) matching would most widen the donor pool.
Open report17% of lung candidates die on the waitlist or are removed too sick before transplant; type-O candidates wait longest (77 d median) and die most — the clinical counterweight to donor-lung discards.
Open reportSign in to browse projects, search the STAR cohort, run a screening on a CT, or mint an SDK token for notebook access.