The United States spends 30% more on health care — and still leaves care undone.

High cost and uneven quality are not an accident; they are the predictable output of how care is organized. Geisinger's response was to treat care delivery as a system that could be re-engineered — not exhorted — toward reliability.

From source systems to the point of care.

One in eight adults. Mostly invisible.

By 2007 the literature was unambiguous: chronic kidney disease affected about 13% of the U.S. adult population¹ — and it behaved like the quietest dangerous disease in medicine. It progressed silently, multiplied cardiovascular risk, and fed a growing dialysis population. At Geisinger, as across the country, it was systematically under-recognized.

Then load the truth into thousands of charts at once.

With a taxonomy worth writing down, we went and wrote it down — everywhere it belonged. The labs already staged every patient; what remained was an engineering problem: get the right staged diagnosis onto the right problem list, safely, at population scale.

The fastest-moving number in the hospital.

No other chronic condition runs on a clock like inpatient glucose. It's measured at the bedside every few hours, swings with every meal, infection, steroid dose, and missed tray — and it's treated with one of the highest-risk medications in medicine. Managing it well is a continuous-control problem. Managing it with yesterday's data is barely managing it at all.

The chart had every number. Nobody was doing the math.

On a typical med-surg unit, vital signs are charted every four to eight hours and land as separate flowsheet rows. Each value gets glanced at on its own. But deterioration rarely announces itself with one dramatic number — it creeps in as several modest ones. The arithmetic that turns five borderline vitals into one alarming patient was being left to whoever happened to look, whenever they happened to look.

Same HL7 lane. Different physiology.

MEWS was the second tenant of the real-time lane built for Glucometrics: vitals charted in Epic stream out through the eGate interface engine, land in the hot partition, and are scored on arrival. No nightly batch, no refresh cycle — the score in the EHR census is the score as of the last observation.

The test worked. The lab worked. The follow-up never happened.

The literature put a number on something every clinician has felt: roughly 8–10% of abnormal laboratory results become "unclosed loops" — the result comes back, and the action it demands never occurs (Singh et al., Am J Med 2010). Not because anyone decided against acting. Because the result arrived between shifts, the letter went unanswered, the patient changed clinics, the chart never resurfaced.

Mine every result. Prove every closure.

Close the Loop is the warehouse pointed at patient safety: enumerate the denominator (every Pap ordered by a Geisinger-employed physician), isolate the abnormals, compute each one's guideline obligation, and search downstream data for evidence the obligation was met. What lacks evidence becomes a worklist — owned by a dedicated CTL team, worked to documented resolution.

Providers told us exactly what was broken. Then the data agreed.

The team surveyed over 250 providers before writing a line of code. The result was a portrait of inversion: the things providers rated most important about documentation were precisely the things the EHR served worst. And a quarter of them were paying for it with more than nine hours a week of non-clinic documentation time.

Built like a product, governed like a clinic.

The team ran SuperNote on four stated principles: best idea wins, anyone can propose ideas, it must be best for patients, and no forced workflows — adoption driven by choice. The artifacts of that culture survive in the slides: design votes, satire, and public self-criticism.

Every gap waited for a visit. Every visit was already full.

The bundle engine could see exactly what each patient needed — sixteen bundles, four to sixteen care needs each, recomputed daily. But seeing isn't doing. Closing a gap still required an office encounter: a nurse to notice the alert, a provider to place the order, a patient who happened to show up. The system's knowledge was continuous; its capacity to act was rationed in fifteen-minute slots.

The business case was sized before it was claimed.

The program quantified its own opportunity the disciplined way: enumerate the population, price every open gap at net revenue, and model capture at 100 / 30 / 20 / 10 percent — presenting the whole ladder rather than the flattering end of it.

3,683 patients filed under a word that means nothing.

"Congestive Heart Failure," ICD-9 428.0 — Geisinger's single largest heart-failure label. It doesn't say systolic or diastolic. It doesn't say why. It doesn't say what good therapy looks like. A diagnosis that can't distinguish a revascularization candidate from a valve-repair candidate isn't clinical information — it's filing. And obesity was worse: the patients weren't mislabeled, they were unlabeled.

From "CHF" to a taxonomy that names the strategy.

The replacement taxonomy factors every heart failure into the two axes that drive management — systolic vs. diastolic, and etiology — eighteen mutually exclusive diagnoses where three generics used to live. Then it meets clinicians where they work: each role sees the slice of the taxonomy it actually uses.

A data-quality problem is a process problem wearing a disguise.

The team's founding axiom: every bad value is the visible symptom of a broken process somewhere in the enterprise — a user interface that demands a phone number and gets (555) 555-5555, a missing organizational definition ("which drugs are beta blockers?"), a column quietly reused for something else. Hunt the anomaly, and it leads you to the thing worth fixing: the UI, the bug, the training gap, the missing standard.

Model the business of care — not the database that recorded it.

The first attempt modeled the data the way the source systems stored it: unwieldy, unnecessarily complex, different for every question. The rebuild went dimensional — conformed patient, provider, and date across every business process — and the payoff was architectural: analytics that survive source-system change, because an abstraction layer shifts that burden to ETL, where it belongs.