Generating Medical Discharge Report Drafts Via Artificial Intelligence (Letter)
A key process in intensive care is the creation of transfer reports. These summarize the essential aspects of the treatment to date and ensure that information is passed on to the next treatment team in a structured manner. They contain, among other things, diagnoses, relevant examination results, and therapeutic measures that have been carried out. The aim of a transfer report is to document the patient's medical history and the course of treatment to date in a concise and complete manner. Incomplete transfer reports can lead to repeated examinations, which can potentially delay further therapy and endanger the patient, as well as incurring additional costs.
In this context, a complete transfer report ensures continuity of treatment and contributes significantly to patient safety.
Transfer reports are created from a synthesis of a large number of documents. In addition to medical facts from tables and free-text summaries, physicians also draw on their experience to pass on relevant information to future treatment teams in the synthesis of detailed documentation. This process naturally requires a great deal of time and concentration in patients receiving intensive care.
Artificial intelligence methods, such as large language models (LLMs), could be used to learn how to create transfer reports and formulate drafts.
LLMs can convert unstructured free text into a structured form, thereby facilitating the report creation process. This would make it possible to generate a pre-formulated draft that doctors could then review and supplement.
At present, transfer reports are created by physicians with considerable effort and time. The methods currently available for automating report generation are limited to text modules that largely bundle data sources deterministically. Some of these methods are already commercially available, but they are not capable of creating individual reports beyond form letters.
The automatic analysis of unstructured text data falls within a subfield of AI known as natural language processing (NLP), which also includes language models (LLMs) such as ChatGPT and Gemini, which have been popular since 2022. A major milestone was reached with the development of the Transformer architecture in 2017, which, in addition to the availability of enormous amounts of training data (e.g., Wikipedia, news articles, or books), forms the basis for many groundbreaking language models. The effectiveness of these LLMs is limited for specific domains such as medicine, as the training data often lacks relevant characteristics such as complex terminology or writing styles. Adapting these models for clinical use to improve patient care, research, and medical education is a highly topical field of research. The most relevant foreign-language models include BioBERT, Meditron, and Med-PaLM. As of July 2024, there is only limited research on the development of German-language models, e.g., medBERT, BioGottBERT, and a few models that are not publicly available.
The sole use of LLMs is problematic due to non-reproducible results, lack of transparency, and bias in the models. Knowledge graphs (KGs) are used in medicine to represent and analyze complex information structures. The creation of a "robust digital patient twin" would be improved by the use of LLMs and KGs. This would allow the state of health to be assessed based on all digitally available clinical information (both structured and, in this project, now also unstructured). Recent studies confirm that adapted LLMs in the English language domain can produce summaries that are equivalent to or better than those produced by human experts, and that these are often difficult to distinguish from reports written by humans.
The legal framework for this results from a complex interaction of legal provisions from data protection law (with regard to health data), artificial intelligence law (AI Act), medical law (in particular hospital law), (medical) product law, medical professional law, and other relevant professional groups (e.g., nursing and pharmacy), as well as relevant civil and criminal law standards of care. In addition, there are fundamental rights requirements based on the ECHR and the EU Charter of Fundamental Rights. Data protection issues relating to the use of health data, in particular for research with AI systems, and professional law issues in connection with automated aggregation and presentation have already been examined in detail (as part of the SMARAGD project, FFG 872575).
調査の概要
状態
条件
研究の種類
入学 (推定)
連絡先と場所
研究連絡先
- 名前:Thomas Tschoellitsch, MD
- 電話番号:78132 +435768083
- メール:thomas.tschoellitsch@jku.at
研究連絡先のバックアップ
- 名前:Jens Meier, MD
- 電話番号:2158 +435768083
- メール:jens.meier@jku.at
研究場所
-
-
Upper Austria
-
Linz、Upper Austria、オーストリア、4020
- 募集
- Johannes Kepler University, Kepler University Hospital
-
コンタクト:
- Thomas Tschoellitsch, MD
- 電話番号:78132 +435768083
- メール:thomas.tschoellitsch@jku.at
-
コンタクト:
- Jens Meier, MD
- 電話番号:2158 +435768083
- メール:jens.meier@jku.at
-
-
参加基準
適格基準
就学可能な年齢
健康ボランティアの受け入れ
受講資格のある性別
サンプリング方法
調査対象母集団
Patients treated in intensive care units.
説明
Inclusion Criteria:
- Treated in intensive care units between 2008-11-01 to 2025-05-31 at the study center.
Exclusion Criteria:
- Refusal to participate in scientific studies.
研究計画
研究はどのように設計されていますか?
デザインの詳細
- 観測モデル:OTHER
- 時間の展望:RETROSPECTIVE
この研究は何を測定していますか?
主要な結果の測定
結果測定 |
メジャーの説明 |
時間枠 |
|---|---|---|
|
Quality of Discharge Report Drafts
時間枠:2008-11-01 to 2025-05-31
|
Quality of Discharge Report Drafts evaluated based on technical parameters and assessed by human experts. Factual correctness: Accurate reproduction of e.g. diagnoses, medication versus hallucinations using Entity Matching Score (Percentage of correctly reproduced entities compared with the source data) and Hallucination Rate (Proportion of unverifiable statements). Completeness: All relevant information is included using checklist compliance (proportion of required fields covered) and Recall against Reference Texts (comparison with physician-written letters, e.g. using ROUGE-L). Clinical Validity: Recommendations must be evidence-based and consistent with clinical guidelines using guideline-adherence (e.g. rule-based systems or LLM-based assessments). Structure is assessed with template compliance using NLP-based section recognition. Data Protection: Degree of Anonymization using NER models. Comprehensibility: Clear, precise language without ambiguities using Readability Assessment. |
2008-11-01 to 2025-05-31
|
二次結果の測定
結果測定 |
メジャーの説明 |
時間枠 |
|---|---|---|
|
Traceability
時間枠:2008-11-01 to 2025-05-31
|
The data sources used to create text passages should be visualized so that the steps performed by the AI can be explained and traced. Traceability should also be evaluated by human experts. |
2008-11-01 to 2025-05-31
|
|
Degree of Necessary Adjustments
時間枠:2008-11-01 to 2025-05-31
|
The degree of necessary adjustments to drafts should be evaluated proportionally as a percentage of the total text and based on the type of adjustments made by human experts (e.g., spelling errors, content corrections, additions). |
2008-11-01 to 2025-05-31
|
協力者と研究者
捜査官
- 主任研究者:Thomas Tschoellitsch, MD、Johannes Kepler University
研究記録日
主要日程の研究
研究開始 (実際)
一次修了 (実際)
研究の完了 (推定)
試験登録日
最初に提出
QC基準を満たした最初の提出物
最初の投稿 (実際)
学習記録の更新
投稿された最後の更新 (実際)
QC基準を満たした最後の更新が送信されました
最終確認日
詳しくは
本研究に関する用語
その他の研究ID番号
- MEDI-DOK Letter
個々の参加者データ (IPD) の計画
個々の参加者データ (IPD) を共有する予定はありますか?
医薬品およびデバイス情報、研究文書
米国FDA規制医薬品の研究
米国FDA規制機器製品の研究
この情報は、Web サイト clinicaltrials.gov から変更なしで直接取得したものです。研究の詳細を変更、削除、または更新するリクエストがある場合は、register@clinicaltrials.gov。 までご連絡ください。 clinicaltrials.gov に変更が加えられるとすぐに、ウェブサイトでも自動的に更新されます。