Medical record summarization for claims: what the summary should show and what a person still reads

Medical record summarization turns a long set of medical records into a short report of the treatment. The summary lists each visit by date, with the provider, the diagnosis and the treatment. Claims adjusters and nurse reviewers use it to understand an injury claim before they read the full file. Each line of a good summary gives the page it came from. AI can draft the summary and check the file, but a person still reads the pages that any decision is based on.

What a claims reviewer needs from a long record#

Say a back-injury claim arrives with 500 pages of records. They come from an emergency room, a family doctor, an orthopedic surgeon, a physical therapy clinic and an imaging center. Some notes are handwritten. The bills sit between the visit notes in no set order.

The adjuster needs a few answers from those pages. What treatment followed the injury? Did the claimant have the same problem before the injury? When can the claimant work again? A useful summary answers from the record itself, one item at a time.

Every item in the summary needs a page number. Use the page number of the compiled PDF. Lawyers often stamp a unique number on every page in a case, called a Bates number. If a law firm has stamped Bates numbers on the pages, use those instead.

Work status can also arrive on a state form. In Texas workers' comp, health care practitioners use DWC Form-073, the Work Status Report, to record a worker's work status and restrictions. A summary should list each of these reports with the dates it covers.

A chronology, a summary and data are three different outputs#

People use "medical summary" for three different outputs. Decide which ones you need before you compare tools.

A medical chronology lists every visit and event in date order. Each entry gives the date, the provider, what happened and the page. It is long, because it is meant to leave nothing out. Legal teams and nurse reviewers use it to check facts.

A narrative summary is a page or two of plain sentences. It tells the story of the injury and the treatment, and it points to the chronology entries that support each statement. Adjusters often start with it. So do the doctors who carry out independent medical exams (IMEs).

Extracted data is the third output. Dates of service, providers, codes and work status dates come out as fields. Software can compare and count fields, and your claims system needs them too. A paragraph of text can't be checked that way.

If you need all three, build the fields first, then the chronology, then the summary. The checks and the chronology are built from the fields. The summary comes last, so it describes a file that has already been checked.

How AI turns a record set into a summary#

The simplest approach sends the whole record set to an AI model and gets a summary back in one step. Nothing in that approach checks the file first. A better approach splits the job into five steps, so each result can be checked. The summary is the last step.

Take the 500 pages from the back-injury claim. First, the software cuts the compiled PDF into separate documents. Then it sorts each one by type. The types in this file include emergency room records, imaging reports, therapy notes, work status reports and bills. Pages it can't sort go to a person. Record review teams call this step indexing. Our healthcare page shows it on a 26-page record, with each document's type, date and provider.

Next, it reads the fields that each type carries. A visit note gives a date of service, a provider and a diagnosis. A bill gives dates, codes and amounts. The software reads handwritten notes too, and it sends a person any field it isn't sure of. OCR for medical records covers this reading step in more detail.

The checks run on those fields before any summary exists. Last, a large language model (LLM) writes the summary from the checked fields and the text of the key documents. An LLM is the kind of AI that reads and writes text. Because it comes last, the summary is written from data that has already passed the checks.

Checks to run before anyone reads the summary#

Each check below compares fields, so software can run it on every file. A reviewer then opens only the checks that failed.

• Every page belongs to the claimantThe name and date of birth on each document match the claim. A page from another patient's chart should never reach the summary.
• Every bill has a visit noteEach billed date of service has a note from the same provider. If it doesn't, either a record is missing or the bill needs a closer look.
• Gaps between visits are flaggedAny stretch longer than the number of days you set goes on a list. The list shows the last visit before the gap and the first visit after it.
• Prior records are markedRecords dated before the date of injury go on their own list, with their pages.
• Work status periods connectEach period starts when the last one ends. A gap or an overlap goes to the adjuster.
• Duplicates count onceThe same visit can arrive twice, once from the provider and once from an attorney. Count it once in the chronology and in the bill totals.
• Nothing is missingPage labels such as "3 of 7" have no gaps in the numbers. For each referral, the file holds the specialist's notes or a request for them.

When a check fails, the reviewer should see the check, the field and the page together. Then the reviewer can fix it without searching the whole PDF. Human-in-the-loop systems covers how to set up that review queue.

Our take. Judge a summarization tool by what it hands your reviewer, not by how well its summary reads. Even when a summary is well written, the nurse still has to check the pages before trusting it. We'd rather the software check the file first and give the adjuster or nurse only the checks that failed, each with its page. Then the summary is the last thing they read, and they already know which lines to question.

Where a person still has to read#

AI summaries make two kinds of mistakes. They add things the record doesn't say, and they leave out things it does. A 2025 study in npj Digital Medicine counted both. All of its authors worked for Tortus AI, and the lead author also lists Guy's and St Thomas' NHS Foundation Trust. They tested their own company's AI system for writing clinical notes.

Clinicians checked 12,999 sentences of clinical notes that an LLM wrote from doctor-patient consultations. In 1.47% of those sentences, they found something the consultation didn't support. Of these errors, 44% were major. A major error could affect a patient's diagnosis or care if no one corrected it. The notes also missed relevant details. Of all the sentences in the consultation transcripts, 3.45% were relevant but missing from the notes. One in ten of the unsupported sentences guessed at the cause of a condition, with nothing in the source to back it.

The study looked at notes written from conversations, not at claim files. But an AI summary of a claim file can make the same mistakes. In an injury claim, the cause of a condition is a question for the adjuster and the doctors, not for the software.

A person should still read the source pages in these cases.

• Any line that a decision depends on, such as a reserve change, a denial, a settlement offer or an IME referral
• Anything a doctor wrote about cause, MMI or impairment
• Records that disagree, such as two providers describing different injuries
• Handwritten notes the software marked as unsure
• Pages the software couldn't sort

No summary decides whether treatment was reasonable and necessary. That's a clinical judgment for your nurse reviewers and physician advisors.

How Docsumo handles a claim's medical records#

The table shows what Docsumo does at each stage.

Docsumo supplies the software, not a review team. Your own adjusters and nurses handle the failed checks and read the pages each decision is based on. If you want a vendor's staff to check every summary, look at services built that way. Enlyte says a clinical expert validates its summaries for bodily injury claims. Wisedocs says human experts with medical backgrounds check its summaries.

Docsumo is SOC 2 Type 2 audited, ISO/IEC 27001:2022 certified, and HIPAA and GDPR compliant. A business associate agreement (BAA) and a data processing agreement (DPA) are available. Your documents are not used to train shared or third-party models. Docsumo runs only in the cloud, and the security page has the details. For the rest of the claim file, see intelligent document processing for insurance and the workflow steps on the platform page.

How to test a summarization tool#

Vendors' accuracy figures come from other teams' files. Test a tool on records your own team has already worked through.

1. Pick three closed claimsChoose claims your team knows well. Include one with handwritten notes, one with a gap in treatment and one with a prior condition.
2. Ask for a page on every lineA summary line with no page number can't be checked quickly, so count it as a miss.
3. Check 20 lines against the pagesPick the lines at random and open each page. Count the lines that are wrong, and the facts the summary left out.
4. Compare the flags with what you knowYour team already found the gaps, missing records and prior conditions in these claims. See whether the tool flags the same ones.
5. Time the reviewNote how long a reviewer takes with the tool's output. Compare it with the time the claim took by hand.

Bring a claim your team has already closed#

A closed claim is the fairest test, because your team already knows the answers. Ask your security team to approve the test first, since the records hold health data. Then read the failed checks before the summary. If the checks find what your reviewers found by hand, the summary is worth reading next.

Book a demo and plan the test around your closed claims, or start a free trial.

Frequently asked questions#