AI in writing and structuring discharge summaries

The discharge summary is both a clinical and an administrative document

The discharge summary, called epikriz in Turkish, is the document the treating physician writes at the end of an inpatient stay or course of treatment. It records the reason for admission, the history and examination findings, test results, diagnoses, procedures and treatments, the patient's condition at discharge, medications and recommendations. At the patient's next visit, the physician who sees them usually learns about the previous stay from this document.

The same document is read again a few steps later for other purposes. The coding team refers to it when coding diagnoses in ICD-10, the billing team when invoicing procedures to SGK, Türkiye's Social Security Institution, and a health insurer when assessing a pre-authorization or claim. The rules for SGK billing are set out in the SUT, the Healthcare Implementation Communiqué published by SGK, and pre-authorization and billing run through Medula, SGK's system. This guide refers to the Turkish system, but the same questions arise wherever discharge summaries feed coding and reimbursement.

Free text is a natural way for physicians to write, but a difficult source for this second use. The same finding appears in different words, abbreviations or Latin terms depending on who wrote it. A secondary diagnosis may stay buried in the text and never reach the coding field, and information copied from earlier notes that no longer applies may find its way into the new document. Across the health system, physicians spend 35% of their working time on documentation, and the discharge summary is part of that burden.

Tasks AI can take on in the discharge summary process

When people think of AI in discharge summaries, they often picture the model writing the text. In practice there are four separate tasks with different risks, and separating them is the first step toward a sound setup.

• Structuring means separating the diagnoses, procedures, medications, key findings and discharge recommendations in the physician's summary into distinct fields. The text itself does not change; each field is linked to the sentence it is based on.
• Code suggestion means suggesting ICD-10 codes from the structured diagnosis fields. Each suggestion comes with the statement and finding it rests on, and a diagnosis mentioned in the text but not coded, or a code with no support in the text, is flagged separately.
• Consistency and completeness checks look at whether the summary contradicts itself or other records. Diagnosis is compared with treatment, test results with their interpretation in the text, and billed procedures with those described; if a report required by SUT is missing, that becomes visible too.
• Drafting means preparing a first version of the discharge summary from the orders, test results, consultations and operative notes created during the stay. It can reduce the documentation burden most directly, but it is also the riskiest task: the model can add information that is not in the record in a fluent sentence, or leave out an important finding.

The first three tasks work on text the physician has written, so their output is easy to link to a source sentence. In a draft, the text itself comes from the model; every sentence should show the record it rests on, and the physician should approve the text.

Steps for using AI in discharge summaries

The following steps apply whichever system is used:

1. Start with structuring, code suggestion and consistency checks; consider drafting only once a workflow that enforces source citation and physician approval is in place.
2. Define an institution-wide field structure for discharge summaries: reason for admission, principal and secondary diagnoses, procedures, key test results, treatment, condition at discharge, discharge medications and recommendations. Physicians keep writing in free text; the fields are filled according to this structure.
3. Require every field and suggestion to be linked to its source sentence, and require what the document says to appear on screen separately from what the system infers.
4. Make ICD-10 suggestions subject to approval; a suggestion should not reach the coding field or the invoice until the physician or coding lead has checked it against its source and approved it.
5. Compare the discharge summary with the invoice and procedure records in both directions: every billed procedure needs a basis in the clinical record, and every procedure described in the summary should appear on the invoice. Complete the reports and documents SUT requires before the invoice is issued, and where possible before the patient is discharged.
6. Record the suggestions the physician corrects or rejects, with the reason; this record shows where the system goes wrong and forms the basis of quality control.
7. Do not enter text containing patient information into tools outside the institution; run the system in the institution's own environment.

Quality control and common error types

When checking the quality of an AI-assisted discharge summary process, it helps to look for these error types separately:

• Fabricated information (hallucination) is a finding, dose, date or procedure that appears in the text but not in the record. Because it is written fluently, it is hard to spot; any sentence whose source cannot be shown should be treated as suspect.
• Omission is an important finding, complication or discharge recommendation dropping out of the summary.
• Negation and uncertainty errors occur when a diagnosis described as "ruled out", "considered unlikely" or "suspected" is taken as a confirmed diagnosis.
• Timing and subject errors occur when a past illness or a diagnosis from the family history is coded as a current diagnosis.

Measure quality on a sample of discharge summaries reviewed jointly by a physician and a coding specialist. Track field extraction accuracy, how many code suggestions are approved or corrected, and sentences whose source cannot be shown, each separately; defining the criteria at the outset shows whether the system improves over time.

General-purpose chatbots such as ChatGPT or Gemini may seem an easy way to edit or summarize a discharge summary. However, these tools are not designed around a specific institution's records, regulations and workflows. Health data is special category personal data under KVKK, Türkiye's personal data protection law, and under the GDPR, so entering patient information into a general cloud-based tool should also be assessed in terms of the institution's responsibility as data controller. We cover the difference in the guide on general-purpose chatbots and clinical decision support systems.

What Opinion AI does in this process

Opinion AI is a clinical decision support platform for hospitals and health insurers, built on MINA, a clinical AI adapted to Turkish. Its hospital platform, TIS, does not replace the hospital's HBYS, the hospital information system; it is added on top of it as an intelligent layer. HBYS and Medula data come together in one live patient profile called the patient twin, and discharge summaries and reports are turned into structured fields within that profile. The physician's text does not change, and each field shows which sentence it is based on.

MINA, working as TIS's clinical assistant, checks each discharge summary for internal consistency, suggests ICD-10 codes and checks SUT compliance. What the document says and what the model infers are shown separately in every answer. Whether a service complies with SUT is decided by the rule engine; AI explains the result and the relevant rule. Medula billing errors and SUT non-compliance are flagged before the invoice is issued, and an ICD-10 suggestion is not final until the physician approves it. The flow is described in detail in our article on structuring the discharge summary.

MINA's open-weight base model has been adapted to Turkish and the clinical language of 14 specialties through continued pre-training (CPT) and specialized by specialty, institution and task with LoRA adapters; knowledge retrieval runs through agentic orchestration over a clinical knowledge graph (GraphRAG). In MINA's latest release evaluation, model results were 87% for ICD-10 coding and 89% for discharge summary field extraction (MINA v3.7 · 8 protocols · ~9,300 cases). A Physician Ethics and Advisory Board of physicians from 15 specialties regularly reviews the model's clinical accuracy and ethical boundaries.

The data stays inside the hospital: TIS is installed on the hospital's own servers or in an isolated cloud environment, the data the model sees passes through KVKK-compliant masking, and the hospital is the data controller. Opinion AI does not collect personal data. People make the decision; MINA shows the rationale and source of its recommendation, and every output can be audited.

Frequently asked questions

Can AI write the discharge summary instead of the physician?

It can prepare a draft, but it should not write the summary in the physician's place. The discharge summary reflects the clinical judgment of the treating physician, while language models can add a finding, dose or date that is not in the record in a fluent sentence, or leave out an important finding. Every sentence in a draft should therefore show its source in the patient's records, and the text should not be finalized until the physician approves it.

What does structuring a discharge summary mean?

Structuring a discharge summary means separating information in the free-text document, such as diagnoses, procedures, medications, key findings and discharge recommendations, into distinct fields that can be searched and checked. The physician's text does not change; the fields sit beside it as a separate layer, and each field is linked to the sentence it is based on. The same information can then be used for coding, billing and the patient's next visit without rereading the text.

Can an ICD-10 code suggested by AI be trusted?

An ICD-10 suggestion is a starting point, not a final code. A good system shows which statement and finding in the discharge summary the suggestion rests on, and separately flags diagnoses mentioned in the text but not coded, and codes with no support in the text. The code should become final only after the physician or coding lead has checked the suggestion against its source and approved it.

Can a general-purpose chatbot be used to write discharge summaries?

It is not advisable for a discharge summary that contains patient information. General-purpose chatbots are not designed around a specific institution's records, regulations and workflows. Health data is also special category personal data under KVKK, Türkiye's personal data protection law, and under the GDPR, so entering patient information into a general cloud-based tool should be assessed in terms of the institution's responsibility as data controller.

Does using AI for discharge summaries take patient data outside the hospital?

It does not have to. Structuring, code suggestion and consistency checks can be done by a system that runs on the hospital's own servers or in an isolated cloud environment dedicated to the hospital, with the hospital remaining the data controller. Opinion AI's hospital platform TIS is set up this way: discharge summaries and reports are processed in the hospital, the data the model sees passes through KVKK-compliant masking, and Opinion AI does not collect personal data.

To assess discharge summary structuring and ICD-10 suggestions on your own discharge summaries, starting with a single module, use the POC Request form. We cover the checks between the discharge summary and the invoice in the guide on preventing SUT non-compliance in SGK hospital billing, and how the discharge summary becomes a single patient profile in the guide on the patient digital twin.