The BRIDGE Protocol
A structured way to reconstruct what actually happened during a hospitalization — and to say how confident we are.
Working draft v0.1 · for teaching LLMs to read medical records
What problem it solves
Most tools that read a medical chart answer “What does the chart say?” BRIDGE answers a harder question: “What really happened to this person, do I believe the chart’s version, and what was never written down?” A discharge summary is a tidy conclusion written under time pressure — it rarely shows how the diagnosis shifted, how sure anyone was, or what question no one asked. The clinical value, and most errors, live in that gap.
The signature move: BRIDGE treats the chart’s silence as data. Its most distinctive output is a list of the specific questions a careful clinician would still ask — surfaced because the record is silent on them — plus a 0–100 certainty score on every conclusion so a reader instantly sees what to trust.
The six stages
BBaseline
Who was this patient before the event?
Example: Knowing Mr. Hanson’s baseline kidney number is what turns an admission value of “3.8” into a real injury rather than “normal for him.” Remove that one prior number and the entire kidney storyline disappears.
RRun-up
What changed in the days-to-weeks before admission?
Example (the classic): a man in liver and kidney failure was a mystery until someone asked what changed — he’d drunk gallons of grapefruit juice while on a high-dose statin. That fact is in no lab or scan. BRIDGE forces the question.
IIn-hospital evolution
How did the diagnosis drift, and what moved it?
Example: Mr. Morrison arrived labeled a “COPD flare.” A chest X-ray showing pneumonia reframed the whole case — changing the main treatment from inhalers to antibiotics.
DDefinitive anchors
Which objective findings actually carry the diagnosis?
Example: Ms. Anderson’s germ was proven by a positive blood culture (certainty ~95). Contrast Mr. Morrison, whose germ came from a weak spit sample (~55). Same disease, very different confidence.
GGestalt at discharge
What was treated, what triggered discharge, and how certain is each diagnosis?
Example: Every diagnosis gets an honest score. A clean-looking list that hides its own uncertainty is a failure — a diagnosis “nailed by a culture” and one that’s a “best-guess label” should not look identical.
EEvaluate & interrogate
Do I agree it’s coherent — and what remains unknown?
Example: Mr. Whitehorse’s real “diagnosis” wasn’t the high potassium — it was the 120-mile trip to dialysis and a broken-down truck. Filed as “noncompliance,” he’d be back in weeks. BRIDGE surfaces the fixable root cause.
The certainty score (0–100)
90–100 Proven — a culture, image, or confirmed measurement
70–89 Strongly supported
40–69 Reasonable, real doubt remains
15–39 Thin — wouldn’t survive a hard question
0–14 The chart is silent
What a BRIDGE reading hands you
Critical data elements
The handful of facts everything hinges on — e.g., a baseline lab value, one X-ray, one culture. If a tool can’t find these, its confidence should drop, not bluff.
What we know
The high-certainty story, stated plainly, each claim carrying its score so a non-specialist sees at a glance what’s solid.
What we don’t know
The explicit questions the chart never answered — the trigger, the missing lab, the social barrier — written as questions a clinician would ask.
Why this matters for building tools: when two independent AI models red-teamed these case write-ups, they caught the same failure again and again — quiet over-confidence and tests that were ordered but never resulted. Those are exactly the slips the certainty score and the “what we don’t know” list are designed to expose. A protocol that makes claims easy to audit is doing its job.