| name | write-report-with-ehr |
| description | Write a context-aware chest X-ray report grounded in the trained MedImageInsight (MI2) adapter AND the patient's EHR record. Adds a COMPARISON section and uses the most recent prior report to call out interval change. |
| argument-hint | - image_path: required path to the current CXR image (PNG or JPEG).
- head: optional adapter head — "mlp" (default) or "svm".
- patient_id: optional override if the image filename does not follow NIH convention.
|
write-report-with-ehr
Three steps: (1) lookup-patient-history, (2) one mcp_medimageinsig_adapter_classify call on the current image, (3) write a comparison-aware report. Self-contained — does not delegate.
Budget
- Tool calls: 1 (adapter, current image only — do not re-score the prior).
- Skill invocations: 1 (
lookup-patient-history).
- Output: ≤ 250 words. No retries on success.
How to use
- Look up the EHR. Invoke
lookup-patient-history with the current image_path. If "Not on file," continue and write COMPARISON: None available.
- Score the current image. Resolve
image_path to absolute (the MCP tool runs in a different cwd). Call mcp_medimageinsig_adapter_classify. Each finding returns positive / possible / negative — already operating-point-bucketed.
- Write the report using the format below, anchoring comparison to the most recent
PRIOR REPORTS block from the lookup.
Why comparison matters
When a prior exists, the clinician already knows what's chronic. Your job is what changed and what to do about it.
- Surface real change — new, worsening, resolved.
- De-escalate chronic findings — one stable sentence, not a paragraph.
- Don't alarm on documented hardware — lines/tubes in
INDWELLING DEVICES are expected; confirm position, do not call them new.
- Anchor recommendations to trajectory — "worsening despite drainage" ≠ "stable post-drainage" even when the imaging looks identical.
Comparison rules
COMPARISON: line is required. Cite modality, view, date, and study_id of the most recent prior report. If none, None available.- Use canonical change words, one per finding: stable, unchanged, improved, worsened, progressed, new, resolved, interval increase/decrease, no significant change.
- Walk Findings in roughly the prior's order. State change (or stability) explicitly.
- Stable findings get one sentence. Save prose for what changed.
- Don't invent priors. If a current finding isn't in the prior report or EHR, write new — not stable.
- Respect the record's last entry. If
LAST RECORD ENTRY NOTE says the record stops before the current study, do not assume what happened in the gap.
- Impression is change-first — interval change first, chronicity last.
Hedging and style
positive → state directly. possible → possible / differential includes. negative → omit, or state if clinically relevant.- State negatives explicitly and grouped. Present tense, impersonal declaratives, plain text.
- The report must not mention MI2, the adapter, classifier scores, EHR JSON, or any model output.
Output format
Return the two trace sections first, then the report block last.
## Prior information used
Bulleted summary of what `lookup-patient-history` returned that you actually
leaned on — HPI, active problems, indwelling devices, the prior study you
anchored COMPARISON to, and any phrases from the prior report you carried
forward. If no prior was available, say so. **Max 6 bullets, one line each.**
## Clinical details used in the report
Bulleted, one bullet per substantive phrase in the report. Each bullet pairs
the report phrase with the source that backed it — an EHR field (`HPI`,
`active_problems`, `indwelling_devices`, `procedures`, `prior report _NNN`)
or an adapter call (`adapter: positive emphysema`). Flag anything not
supported by either source. **Max 10 bullets, one line each.**
<confidence>{confidence}</confidence>
<report>
INDICATION: {one sentence using the admission HPI / current clinical question}
TECHNIQUE: {one sentence (view, technique)}
COMPARISON: {modality view, date, study_id of most recent prior — or "None available."}
FINDINGS: {impersonal declarative sentences in prose paragraphs. Each finding tagged with a canonical change word. Confirm positions of documented indwelling devices.}
IMPRESSION: {change-first prose, one short paragraph or numbered "1. ... 2. ..." inline on one line — no markdown lists.}
</report>
The <report> block is plain text only — no markdown. No bullets, no -/* lists, no numbered-list line breaks, no bold/italic, no headers. Section labels (INDICATION: etc.) are the only structure.
{confidence} is low / medium / high, derived from the adapter mix and image quality.
