Lumen in healthcare. The platform underneath →
Lumen Health Partners provides the connective tissue between healthcare systems — intelligence that moves through them, not around them. We keep the source beneath the models you already run, so the evidence is still reachable when the question changes.
None of it would need building if that were already true.
Every organization has a list of questions it stopped asking. Not because they didn't matter — because answering one properly meant four datasets, a stratification nobody anticipated, and a metric that wasn't in any report. A week of work for an answer the meeting had already moved past.
Two things put it out of reach. Some of the evidence didn't survive the trip in. Reaching the rest cost more than the answer was worth.
The answer was already there. Nobody could afford to look.
cohort shrinkage from cumulative ETL steps when full data exhaustivity is required
Priou et al., 2024per-step ETL transfer rates that look acceptable in isolation
Priou et al., 2024the dominant missingness pattern in EHR data — non-random and clinically informative
Haneuse et al.; multipleEvent ordering and timing flatten into date fields.
"Rule out", "suspected", "family history" don’t survive coding.
Who recorded it, why, and from where, drops off in transit.
Labs ↔ diagnoses ↔ outcomes break apart in flight.
Schemas forget. Lumen remembers.
Priou S, et al. "Where have my patients gone?" — A simulation study on real-world data processing in Clinical Data Warehouses. ScienceDirect, 2024.
Haneuse S, Daniels M. A General Framework for Considering Selection Bias in EHR-Based Studies. Am J Epidemiol, 2016.
Klann JG, et al. The Generalized Data Model for Clinical Research. BMC Med Inform Decis Mak, 2019.
The cost has a cause, and it was built in.
For thirty years, healthcare has shoehorned reality into rigid schemas. ETL, ELT, early-binding, late-binding, "right-time" binding — every one of them decides what to carry and drops the rest. The unstructured note. The free-text exception. The signal that didn't have a column.
The problem was never the model. It was that the model became the only copy. Once the source is gone, every step after it can only lose more, and no later question can reach what an earlier one discarded.
So keep the source. Then project as many models over it as the work requires — curated, governed, fast. When the question changes, re-project instead of re-engineer.
Data must conform. What doesn't fit, dies in transit.
Per-step transfer rates often look fine — Priou et al. measured 73–100% per step. Multi-step cumulative loss is what hurts.
Views are built over a source that is kept. When the question moves, the view is rebuilt from it.
Continuity preserves the narrative nuance, temporal context, and provenance that a pipeline loses when the source does not survive it.
Illustrative — cumulative loss compounded across a six-step pipeline, from the per-step rates cited above.
Keeping the source only helps if three things hold.
Clinical notes, lab results, images, claims — everything is preserved in the form it arrived in. Models are projections over that record, not replacements for it. No column, no row, no enum can quietly make a signal unrecoverable.
The same record can be irrelevant today and decisive tomorrow. We re-evaluate against the current question, not against a frozen schema written six years ago.
Embedded intelligence joins, weighs, and synthesizes — turning the lake of preserved signal into clinical and operational answers, fast enough to act on.
Curated views are how access limits and query performance get expressed. Continuity doesn't remove them — it puts something underneath them, so a view can be narrowed, rebuilt, or replaced without another migration. Nothing is migrated to adopt it. The warehouse stays where it is; the source goes underneath.
Continuity within boundaries. Keeping the source raises the question of who can reach it.
That's half of it. The other half is what a question can do once the source stays put.
A dashboard answers the questions it was built for. It cannot answer one whose shape depends on a finding not yet made — where the second question is set by the first answer, where ruling something out means reaching across domains that were never modelled together, and where the metric that finally settles it was in no report.
That loop is what a preserved record makes possible, and it is what turns a week of analyst work into minutes. Lumen sits between the systems you have and the answers you need.
Grain that double-counts when summed, maturity windows before a rate means anything, reversals that have to be excluded — these are built into how the query is constructed and shown in the result. Prompt instructions are not controls.
An answer arrives with the records behind it, so the person asking can check it rather than trust it. A number nobody can trace is a number nobody should act on.
Below, one of those loops — run end to end.
A real investigation, run on a client's live operations. Management arrived with two movements in volume that looked like they needed explaining. The work took minutes. What it returned was not what anyone went looking for.
Two apparent anomalies in volume, both surfaced by leadership, both looking like business events that needed a cause. Explain them.
Neither was a business event. Both resolved to scheduling artifacts — the calendar moving underneath the measurement, not the work changing. Two questions answered and closed.
A third movement was sitting in the same data: a step change a couple of weeks old that no one owned and no report had flagged. Isolating it meant stratifying on a variance structure that only became visible mid-analysis, ruling out three hypotheses against three different datasets, and finally settling it on a lag metric that appears in no standing report.
Every step was chosen by the one before it. The stratification did not exist until the second answer suggested it. The third dataset was out of scope until the second hypothesis failed. The metric that settled it had never been built, because until that morning nobody had needed it.
A dashboard could not have produced this. Nobody knew to build it.
What that investigation did once, these were built to do on demand.
Five are in production. Connect is in development and marked as such. All of them run on your organization's own preserved record and the same investigative loop — reasoning grounded in the record it runs on, and a team that has done the work. They also work in concert rather than in sequence: a finding in one reaches the others as it happens. The technology is the easy part.
Allocates cost against clinical activity — prescriptions administered, nursing visits, patient days — instead of square footage. Every dollar traces to the driver that placed it.
Reads the documents a claim generates and ranks the open work by what a delay actually costs — so the work that quietly becomes next month's crisis gets picked today.
Queries the continuous record rather than the warehouse, so a population query returns the cohort that was always there.
Works an authorization from the chart rather than the claim — the visit notes, the imaging, the alternatives already tried.
Designed to give a patient a navigator that knows their record — visits, medications, recent labs, language — deployed under your brand.
Surfaces what matters at the moment of decision — prior conversations, social determinants, family history — while drafting the documentation that follows.
Nothing in the loop is specific to a department. A preserved record and a question that moves work the same way against pharmacy billing as against surgical scheduling — the domain changes what the answer is about, not how it gets found.
That is the claim, and the range below is the evidence for it. Each of these is work the same platform has already been pointed at.
Same platform each time, implemented for the organization in front of it — which is the subject of the section below.
Lumen builds the technology once and implements it inside your environment under contract. That is the model, and it is what the range above is made of — the same platform each time, fitted to the organization in front of it. And where an organization would rather build its own solution on that platform, it can — on the same preserved record, inside the same boundaries.
Healthcare technology gets adopted at the speed of trust, not the speed of features. We embed alongside your team — sharing accountability for outcomes, not just delivery. Our work spans implementation, change management, and the slow craft of getting AI to behave well in front of clinicians.
Engineers and clinicians on-site, not on a Zoom from a thousand miles away.
Pricing tied to measurable results — not seat licenses or API calls.
Your operational realities shape the product. The product's capability shapes your operations.
That is the argument. What follows is only about the name.
In anatomy, the lumen is the open passageway inside a vessel — the space that lets blood, air, light pass through unobstructed. It's also the unit of luminous flux: how much light a source actually emits into the world. Our name carries both.
Healthcare data is meant to move. Between provider and patient, between record and decision, between insight and intervention. We build the lumen — the open passage that lets it flow without friction, without loss, without translation tax.
Four arcs facing inward. Patients, providers, payers, and platforms — all of them oriented toward a shared center. We're called Partners because no one fixes healthcare alone. The work happens at the meeting point.
That is the name. The rest is a conversation.
We're selective about partnerships — we work best with health systems, payers, and digital health companies ready to do the integration work that makes AI actually useful. Tell us where you're stuck.