How to summarize a medical record, step by step

By Ahmed Jemaa, Co-Founder and CEO of Medrecords AI. Updated 2026-11-12.

Limits: We have not timed the listed tools on a shared test file: time figures come from the cited studies and vendor pages, and the worked example is synthetic.

This is the method for a physician or nurse reviewer who reads the file and stands behind the summary. It works the same whether you do steps 4 to 6 by hand, in a template or with software. Each step ends with an output you can point to later.

Step 1: write the question before you open the file

The question decides what goes in and what stays out. Without it, a summary drifts into retelling every visit.

• IME or QME: copy the referral questions word for word into a numbered list. Typical ones: diagnosis, causation, whether treatment was reasonable, MMI, impairment, work capacity, future care.
• Peer review: write the treatment under review, the dates of service and the criteria or guideline you will judge it against.
• Referral: write the single question you want the consultant to answer.
• Transfer: write what the receiving team must do in the next 24 hours.

Output: a question list at the top of your working file.

Step 2: number every page

Apply Bates numbers, or at least a running page number, before you read. Every fact you extract will carry this number, and every later check depends on it. Adobe Acrobat Pro has a built-in Bates numbering tool (Adobe help, read 2026-11-12). If records arrive in several PDFs, merge them in the order received first, so page 1 is always the same page.

Output: 1 numbered file, plus a note of which sender supplied which page range.

Step 3: remove duplicates

Record sets assembled from several senders repeat the same documents: the ED note from the hospital and again from the PCP, the MRI report from radiology and again in the surgeon's packet. Reading copies wastes time, and copies with small differences (a fax header, an addendum) can hide a late entry.

• Mark exact copies and keep the first instance.
• Compare near copies side by side. A copy with an added addendum or a different signature date is not a duplicate; it is evidence of a change.
• Log what you removed by page number, so you can show it later.

Output: a duplicate log (kept page, removed pages, why).

Step 4: sort and index

Sort by date of service, not by sender. Then skim each document just long enough to list it: date, provider or facility, document type, page range. Don't extract yet. The index shows you the shape of the case (when care started, who treated, where the gaps are) before you read in depth.

The index also does 2 jobs later. It becomes the "records reviewed" section of an IME report, and it is where you spot missing records: a referral to a surgeon with no surgeon's note, an MRI ordered with no report.

Output: a records index, 1 line per document.

Step 5: extract each encounter with its page

This is where the hours go. For each encounter that bears on your question, record the same fields every time:

2 rules prevent most later errors. Quote, don't paraphrase, anything that bears on causation, side or severity ("L5-S1 left paracentral extrusion contacting the left S1 root", not "disc problem"). And keep pertinent negatives: "SLR negative" on 1 date is how you show improvement.

The BMC study of manual chart abstraction found a median of 60 seconds per imaging report to pull a few fields (BMC Med Inform Decis Mak 2022). A file with 60 encounters at that rate is an hour of extraction on top of reading.

Output: an extraction grid, 1 row per encounter. This is your chronology; the templates page has a copyable grid.

Step 6: write the summary for its reader

Write from the grid, not from memory. Use the template for the purpose and lead with what answers the question.

• Start with 2 or 3 sentences that answer the question as far as the records allow. From the synthetic worked example: "Records document a lifting injury on 2025-03-04 with an L5-S1 disc extrusion contacting the left S1 root on MRI, treated conservatively with 1 epidural injection; full duty from 2025-08-18."
• Then the supporting history, grouped by phase (pre-injury, acute, treatment, current status) with a page cite after each fact.
• Keep the records' words for diagnoses and findings. Your opinions go in a separate, labeled section.
• Cut what doesn't serve the question. A billing ledger, cover letters and repeated work status forms can be 1 line each.

Output: a summary draft with a page cite on every fact.

Step 7: flag what is missing or in conflict

Readers turn to this section first, and AI drafts miss it most often. Check the file against itself for:

• Referenced but absent: tests ordered with no result, referrals with no consult note, an "outside records reviewed" with no outside records.
• Gaps in care: periods with no treatment, and whether a note explains them.
• Conflicts: left vs right, dates that don't line up, a diagnosis that changes without a new finding.
• Prior history of the same body part: often in a PCP note or pharmacy printout years before the event, and decisive for causation and apportionment.

Output: a flag list, each with page numbers, and a records request if something is missing.

Step 8: verify, then sign

Verification is not rereading the file. It is checking the lines you will rely on:

1. Open the cited page for every fact an opinion rests on and confirm date, side, value and wording.
2. Spot-check at least 10 other lines at random. If 1 fails, check the rest of that section.
3. Read in full any record your opinion depends on, not only the cited line.
4. Record in the report which records you reviewed and, if software or another person excerpted them, say so. California requires the physician to review such excerpts (Labor Code 4628(c)).

The full verification list, with a score, is on the quality checks page.

Output: the signed summary, plus the index, grid, duplicate log and flag list as your working file.

Questions about the steps