The BRIDGE Protocol — Process Flow Sheet

A step-by-step view for mapping your extraction requirements onto a clinical-reasoning process. Each step shows the question it asks (its input) and the question it answers (its output), with a worked example and the kind of IT requirement it maps to.

For technical readers · not clinically technical

The business case in one idea

A discharge summary reliably records the proximal cause of a hospitalization — the thing that happened last. It rarely records the root cause — the change in the patient's life that set the proximal cause in motion. A system that captures only the proximal cause produces a tidy record and a patient who is readmitted next month. A system that surfaces the root cause routes to the intervention that actually prevents the next admission. That gap is where the money and the risk live — in readmission penalties, risk stratification, and care-management targeting. BRIDGE is a repeatable way to make that gap visible, and to say — with a 0–100 confidence score — how sure it is at each step.

The worked example: the pork-rind patient

What the chart says (proximal)

“Dietary indiscretion”

A dialysis patient arrives fluid-overloaded 12 hours before their scheduled treatment. The salt in a bag of pork rinds drove the volume overload. Accurate — and useless for preventing the next one.

What BRIDGE surfaces (root)

A life that changed

The spouse recently died; the patient can no longer drive; between buses he walks to the corner store and buys the cheap, salty food he can afford. Something changed — and it will keep changing the outcome until someone asks. That is a social-work referral, not a diet lecture.

The six steps — what each asks and answers

BBaselineestablishes “normal for this patient”
The question it ASKS

Who is this patient at steady state — their usual labs, weight, medications, and routine?

The question it ANSWERS

What counts as abnormal for this person — turning a raw value into a signal.

Pork-rind case: his usual between-session weight gain, so today's excess is measurable.  From our cases: Mr. Hanson's baseline kidney number is what makes an admission value an injury rather than “normal for him.”
Maps to requirements likepatient-specific baselinesreference-range normalizationrecord linkage / entity resolutionlongitudinal history ingest
RRun-upfinds the proximal trigger — and its absence
The question it ASKS

What changed in the days-to-weeks before admission — new meds, foods, missed treatments, life events?

The question it ANSWERS

The proximal trigger — and whether the record even contains it. A silence here is itself a finding.

Pork-rind case: the salt load is the proximal trigger the chart can hold; the bereavement and lost transport are the ones it usually doesn't.  From our cases: the classic grapefruit-juice example, or Mr. Morrison's “cold last week.”
Maps to requirements liketemporal windowingchange detectionmedication reconciliationSDOH / outpatient note ingestmissing-data flags
IIn-hospital evolutiontracks the diagnosis as it moves
The question it ASKS

How did the working diagnosis shift during the stay, and what piece of evidence moved it?

The question it ANSWERS

The diagnostic trajectory — not just the endpoint — and who/what changed it.

Pork-rind case: “fluid overload” confirmed quickly; little drift.  From our cases: Mr. Morrison arrived a “COPD flare” and a chest X-ray reframed him to pneumonia — changing the treatment.
Maps to requirements likeevent / timeline modelingnote-type sequencinghypothesis trackingprovenance per claim
DDefinitive anchorsseparates proof from assumption
The question it ASKS

Which objective findings actually prove the diagnosis — a culture, an image, a trended lab?

The question it ANSWERS

What is load-bearing vs. merely assumed — and how confident to be (the certainty score).

Pork-rind case: weight above dry weight + the treatment response anchor it.  From our cases: Ms. Anderson's germ was proven in blood culture (~95); Mr. Morrison's came from a weak spit sample (~55) — same kind of disease, very different confidence.
Maps to requirements likeevidence gradingstructured-vs-narrative extractionconfidence scoringclaim-to-source linking
GGestalt at dischargeaudits the discharge picture
The question it ASKS

What was treated, what triggered discharge, and how certain is each diagnosis on the list?

The question it ANSWERS

An auditable discharge picture with a calibrated confidence on every diagnosis — no tidy list that hides its own doubt.

Pork-rind case: dialysis fixed it; the summary says “dietary indiscretion” — correct, proximal, and the trap.  From our cases: a diagnosis clinched by a culture and one that's a best-guess label should never look identical.
Maps to requirements likedischarge-summary parsingprincipal vs. secondary dxcertainty taggingcoded output (ICD/SNOMED)
EEvaluate & interrogatethe root-cause step
The question it ASKS

Do I believe this story is coherent — and what did the chart never say that I'd need to ask?

The question it ANSWERS

A coherence check plus a ranked list of the unasked questions — where the root cause lives.

Pork-rind case: this is the step that asks “why the salty food, why now?” and surfaces the bereavement + lost transport — turning “noncompliance” into a fixable problem.  From our cases: Mr. Whitehorse's real diagnosis wasn't the potassium — it was the 120-mile trip to dialysis and a broken-down truck.
Maps to requirements likeconsistency / contradiction detectionSDOH gap detection“known-unknowns” generationcare-management / referral triggers

The confidence score every step carries (0–100)

90–100  Proven — culture, image, confirmed value
70–89  Strongly supported
40–69  Reasonable, real doubt
15–39  Thin
0–14  The chart is silent

Mapping your requirements to the process

If your requirement is…Step(s)…this is what to build there
“Normalize a lab/value against the patient's own history, not a population range”BPatient-specific baselines with record linkage across prior encounters.
“Detect what changed before the admission (meds, diet, missed care, life events)”RTemporal windowing + change detection; ingest outpatient/ED/SDOH notes; flag when the trigger is absent.
“Show how the diagnosis evolved, with an audit trail”ITimeline/event model with per-claim provenance and note sequencing.
“Attach a confidence level and cite the evidence for each finding”D · GEvidence grading + calibrated confidence scoring, each claim linked to its source line.
“Parse the discharge summary into principal vs. secondary diagnoses”GDischarge-summary parser producing coded, certainty-tagged output.
“Catch contradictions the chart didn't reconcile”EConsistency/contradiction detection across notes (e.g., conflicting values or meds).
“Flag missing social-determinant / root-cause information and trigger a referral”ESDOH gap detection + a “known-unknowns” generator that emits questions and routes to care management.
“Reduce 30-day readmissions / improve risk stratification”R · EThe proximal-vs-root distinction is the lever — Run-up finds the trigger, Evaluate finds why it happened.
How to use this: put your requirements list beside the mapping table, find the step that owns each one, and read that step's asks/answers to see what the input and output actually are. The steps run in order, but a requirement usually lives in one or two of them.  ·  Note: all patient examples are fully synthetic test data, not real patients.  ·  See the companion BRIDGE_One_Page_Summary.html and the framework spec in this folder; the project's output & indexing policy is in top-level CLAUDE.md.