Discharge summary extraction turns the hospital's closing note for a stay into named fields: admission and discharge dates, discharge diagnoses, procedures, a medication list split by status, and follow-up appointments and instructions. Each value is copied as written, tied to the page it came from, and checked by a person when the software is unsure of it.
What does a discharge summary contain as data?
A discharge summary is the document a hospital writes when a patient leaves, and as data it holds a header, a set of diagnoses, a list of procedures, a hospital course narrative, a medication list and follow-up instructions. Some of those parts are short lists that become clean fields. The hospital course is prose that mostly stays prose.
Two published sources describe what belongs in the document. The federal hospital rule in 42 CFR 482.24 on medical record services requires a "discharge summary with outcome of hospitalization, disposition of case, and provisions for follow-up care". The HL7 C-CDA on FHIR Discharge Summary profile goes further for electronic documents. It requires three sections: Allergies and Intolerances, Hospital Course and Discharge Diagnosis. It lists many optional ones, including Discharge Medications, Procedures, Hospital Discharge Instructions, Hospital Consultations and Plan of Treatment.
A scanned summary from a real hospital rarely follows that list exactly. The headings vary, some sections are merged, and dictated summaries may have no headings at all. The profile is still a good checklist for what to look for, because most summaries cover most of those sections under some name.
A field map of one fictional discharge summary
Every value below is made up for the example. The patient, the record number and the dates are fictional.
| Section | Field | Example value |
|---|---|---|
| Header | Patient name | Rosa Delgado-Finch (fictional) |
| Header | Medical record number | EX-4471209 |
| Header | Admission date | 2026-03-02 |
| Header | Discharge date | 2026-03-05 |
| Header | Attending physician | Dr. Alan Pryce (fictional) |
| Header | Discharge disposition | Home with home health |
| Diagnoses | Principal diagnosis (as written) | "Community-acquired pneumonia, right lower lobe" |
| Diagnoses | Secondary diagnosis 1 (as written) | "Type 2 diabetes mellitus" |
| Procedures | Procedure 1 | "Chest X-ray, two views", 2026-03-02 |
| Allergies | Allergen and reaction | Penicillin, rash |
| Medications | Entry 1 | Example drug A 500 mg by mouth twice daily, status: new |
| Medications | Entry 2 | Example drug B 10 mg by mouth daily, status: stopped |
| Follow-up | Appointment | Primary care, Dr. Lena Ortiz (fictional), within 7 days |
| Follow-up | Pending result | Blood culture final report, to primary care |
| Signature | Signed by and date | Dr. Alan Pryce, 2026-03-06 |
The map records what the document says, not what it means. Extraction copies the diagnosis text and the medication lines as written. Interpreting them is clinical work, and coding them is the coder's work.
Discharge diagnoses and procedures as fields
Discharge diagnoses and procedures should be captured as ordered lists, one entry per line, with the text exactly as written, the order it appears in, and any code printed beside it. The first-listed diagnosis is usually labelled principal or primary. The rest are secondary. Keeping the printed order lets a coder or reviewer see the physician's own ranking.
The terms come from coding rules, not from the summary itself. The FY 2026 ICD-10-CM Official Guidelines for Coding and Reporting quote the Uniform Hospital Discharge Data Set definition of principal diagnosis as the "condition established after study to be chiefly responsible for occasioning the admission of the patient to the hospital for care". The same guidelines define other diagnoses as conditions that coexist at admission, develop later, or affect treatment or length of stay.
That definition belongs to the coder applying it. An extraction step should not decide which condition is principal. It should record which line the summary labels principal, and leave a blank label when the summary does not say.
Useful fields for each diagnosis entry:
- Position, such as 1, 2, 3, in printed order.
- Label, such as principal, secondary, or none, copied from the document.
- Text, the words exactly as written.
- Printed code, when the summary shows one beside the text, kept separate from the text.
- Source, the page and the spot on the page.
Diagnoses that appear outside the diagnosis list
A discharge summary often names conditions in the hospital course narrative that never make it into the numbered diagnosis list. A sentence such as "a mild rash was noted on day two and resolved" mentions something the list does not. Do not move narrative mentions into the diagnosis list. The list is what the physician chose to report as diagnoses, and adding to it changes the document's meaning. If a project needs narrative mentions, capture them in a separate field with its own label, such as "conditions mentioned in hospital course", and route them to review.
The reverse also happens. A diagnosis list can carry a line the narrative never explains. That is not an extraction problem, so record the line as written and move on.
Procedure entries follow the same pattern, with a date field added. Many summaries list procedures without dates, so the date field needs a "not stated" value rather than a guess.
Where ICD-10 and CPT codes show up in clinical documents, and why a printed code is not always the code that was billed, is covered in the post on where ICD-10 and CPT codes appear in clinical documents.
Discharge medication lists: new, changed and stopped
A discharge medication list should be split into one entry per drug, with a status field that says whether the drug is new, changed, continued or stopped. Many summaries group drugs under those headings. Others mark each line with a word such as "NEW" or "STOP". Capturing the status is what makes the list useful later.
The grouping is where extraction goes wrong most often. A summary may print four headings: "Start taking", "Change how you take", "Continue taking" and "Stop taking". If the extraction step flattens all four into one list, a stopped drug looks exactly like a current one. That error is invisible in the extracted data unless the status was kept.
For each entry, keep the status as its own field, and keep the original heading text beside it, so a reviewer can see why the status was assigned. When a changed drug shows both an old and a new dose, store both, labelled, rather than only the new one.
A fictional example of how one section splits:
| Printed heading | Printed line | Status field |
|---|---|---|
| Start taking | Example drug A 500 mg twice daily | New |
| Change how you take | Example drug C, was 20 mg daily, now 10 mg daily | Changed (old: 20 mg daily, new: 10 mg daily) |
| Continue taking | Example drug D 1 tablet daily | Continued |
| Stop taking | Example drug B 10 mg daily | Stopped |
The drug names here are placeholders. The point is the structure, not the drugs.
The per-drug fields themselves, such as strength, route, frequency and prescriber, and the paperwork behind medication reconciliation, are covered in the guide to the fields every medication list entry needs.
How should follow-up instructions in a discharge summary be captured?
Follow-up instructions in a discharge summary should be captured as separate entries by type: appointments, pending tests or results, and written instructions. An appointment entry holds who, when and where. A pending item holds what is pending and who will receive it. Instructions are kept as text, because their wording matters.
Follow-up is part of the federal definition of the document. The 482.24 rule names "provisions for follow-up care" as one of the three things a discharge summary must cover. The discharge planning rule, 42 CFR 482.43, also requires the hospital to send necessary medical information to the providers responsible for the patient's follow-up care at the time of discharge.
The fields worth capturing for an appointment:
- Provider or clinic, as named.
- Timing, kept as written. "Within 7 days" is not a date, and turning it into one is a guess.
- Scheduled date and time, only when the summary prints one.
- Reason, when stated.
For pending items, record the test name as written, the stated recipient, and any phone number or instruction attached. For written instructions, such as activity or diet notes, keep the full text in one field and do not split it into pieces that lose their meaning.
The common mistake is converting relative timing into a calendar date. "Follow up in 1 to 2 weeks" becomes 2026-03-12 in one system and 2026-03-19 in another. Keep the text, and add a computed date only if the receiving system needs one, labelled as computed.
Discharge summary layouts that break extraction
Discharge summary layouts break extraction most often through dictated summaries with no headings, addenda that change earlier content, and medication tables that run across page breaks. Buried summaries, pen corrections and outside templates add more. Each one causes values to land in the wrong field or disappear, often without any visible error in the extracted result.
Dictated summaries. A transcribed dictation can read as one long block of text, with sections announced in speech ("Discharge medications are as follows") rather than printed as headings. Section detection then depends on phrases, not layout. Fields from these documents are good candidates for review, because a missed section boundary moves a whole list.
Addenda. A summary may be signed, then amended with an addendum days later. The addendum might correct a diagnosis or add a pending result. If extraction reads only the first document, the corrected value is lost. Keep each addendum as its own document, with its own signature date, linked to the original.
Multi-page tables. A medication table that starts on page 3 and continues on page 4 may repeat its header row, or may not. Rows can split mid-line, with the drug on one page and the dose on the next. A row that appears to have no dose is often a row cut by a page break.
Summaries buried in a longer record. A discharge summary that arrives inside a 300-page records release has to be found before it can be read. Its first page may look like a progress note, and its last page may run straight into the next document. Record the first and last page of the summary as fields of their own, so every extracted value can be traced to the right document and not to a neighbour.
Handwritten changes on a printed summary. A printed summary with a pen correction, such as a crossed-out dose with a new one written beside it, holds two values in one spot. Extract both, flag the line, and let a person decide which one the document now says.
Templates from outside hospitals. A practice that receives summaries from several hospitals sees several layouts for the same content. The fields stay the same. Where they sit on the page changes.
Every field from these layouts should carry its page location, so a reviewer can check it against the original without hunting. Schema Care reads discharge summaries for diagnoses, procedures, medications and follow-up instructions, scores each field, and routes low-confidence ones to a reviewer; the full list of supported paperwork is on the documents Schema Care reads section. Summaries from other facilities often sit inside longer charts, which the guide to clinical document extraction sorts by type, and the steps for any record are in how to extract data from medical records. Before real summaries reach any tool, read the HIPAA and PHI category on what has to be in place.