このページは自動翻訳されたものであり、翻訳の正確性は保証されていません。を参照してください。 英語版 ソーステキスト用。

Shadow-Mode Evaluation of an AI Assistant for Multidisciplinary Tumor Boards (OnkoBot)

2026年11月9日 更新者:Lisa Adams、Technical University of Munich

OnkoBot: Prospective Shadow-Mode Evaluation of a Locally Operated, Source-Grounded Large Language Model System for Drafting Multidisciplinary Tumor Board Recommendations

This study tests whether an artificial intelligence (AI) system that drafts tumor board recommendations performs as well on new cases as it did on archived cases. The system runs in shadow mode at three multidisciplinary tumor boards of TUM University Hospital rechts der Isar. For each case it drafts a recommendation before the board meets. The board does not see the draft, and the draft has no influence on patient care. After the meeting, each draft is compared with the board decision. The agreement rate is then compared with the rate measured on archived decisions of the same boards.

調査の概要

状態

募集

条件

介入・治療

詳細な説明

The system is a locally operated large language model (LLM) pipeline with retrieval-augmented generation. Its input is the case as presented to the board, with diagnosis, course, question and all patient documents dated before the board date. It retrieves guideline recommendations (Onkopedia and AWMF S3 guidelines in the version in force on the board date), laboratory values, precedent cases from the board archive and PubMed abstracts published before the board date. It then drafts a recommendation with justification. Each statement of the draft is linked to a verbatim quote from its source. All models run on local GPU servers of the hospital with no data retention beyond the immediate analysis, and no patient data are sent to external providers.

The study includes the hematology board, the orthopedics and sarcoma board and the head and neck board. For each case registered for a session, the system drafts a recommendation after the pre-board documents are pulled and before the session starts. Board members do not see the drafts. The decision pipeline (generator Gemma-4-31B at temperature 0) is frozen and versioned for the study period. The study uses routine clinical data only. Patients are not contacted, and no study-specific procedures are performed.

The draft is compared with the consensus decision in the board report. An LLM judge (GLM-5.3-Flash, majority of three votes) assigns both texts to one of seven decision categories: systemic therapy, surgery, radiotherapy, watch and wait, further diagnostics, best supportive care and referral to the molecular tumor board. It also scores agreement in substance from 0 to 3 and records the direction of any disagreement. The judge was validated against clinician ratings on a sealed test set. The same prompt version is used to score the prospective cases and the retrospective reference sample, separately from any output that clinicians see.

The retrospective reference is the evaluation sample of archived decisions of the same three boards (about 2,000 cases, one per patient, drawn with a fixed seed and stratified by time). Patients who also appear in the prospective cohort are removed from the reference. The primary analysis tests non-inferiority of the prospective category match rate against the retrospective rate, pooled over the three boards, with a margin of 7.5 percentage points and a one-sided alpha of 0.025. The variance is cluster-robust for board sessions. A case without a draft counts as not concordant. Per-board results are reported descriptively with 95% confidence intervals. An exploratory component evaluates a trial-matching engine that proposes recruiting clinical trials per case.

In a subset of 30 to 50 cases per board, physicians rate the system draft against the board decision. This subset tests whether the agreement between judge and clinicians, measured retrospectively, also holds prospectively.

研究の種類

観察的

入学 (推定)

450

連絡先と場所

このセクションには、調査を実施する担当者の連絡先の詳細と、この調査が実施されている場所に関する情報が記載されています。

研究連絡先

  • 名前:Lisa Adams, MD
  • 電話番号:+49 89 4150 1084
  • メール:lisa.adams@tum.de

研究連絡先のバックアップ

研究場所

    • Bavaria
      • Munich、Bavaria、ドイツ、81675
        • まだ募集していません
        • TUM University Hospital rechts der Isar
      • Munich、Bavaria、ドイツ
        • 募集
        • TUM Klinikum rechts der Isar
        • コンタクト:

参加基準

研究者は、適格基準と呼ばれる特定の説明に適合する人を探します。これらの基準のいくつかの例は、人の一般的な健康状態または以前の治療です。

適格基準

就学可能な年齢

  • 子
  • 大人
  • 高齢者

健康ボランティアの受け入れ

いいえ

受講資格のある性別

全て

サンプリング方法

非確率サンプル

調査対象母集団

All cases registered for the hematology, orthopedics and sarcoma, and head and neck tumor boards of TUM University Hospital rechts der Isar during the study period. One case is one patient presentation on one board date.

説明

Inclusion Criteria:

  • Case registered for a session of a participating tumor board (hematology, orthopedics and sarcoma, head and neck) of TUM University Hospital rechts der Isar during the study period
  • Board registration and pre-board patient documents available before the session starts

Exclusion Criteria:

  • Case not discussed at the session, defined as no board report with a consensus decision for that date (deferred or withdrawn registrations)
  • Session without a shadow run of the system

Excluded cases are counted and reported by reason.

研究計画

このセクションでは、研究がどのように設計され、研究が何を測定しているかなど、研究計画の詳細を提供します。

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:COHORT
  • 時間の展望:PROSPECTIVE

コホートと介入

グループ/コホート
介入・治療
Hematology board

Consecutive cases presented at the hematology tumor board of TUM University Hospital rechts der Isar.

A locally operated, source-grounded large language model system drafts a tumor board recommendation with justification and verified citations for each case before the board meeting. The draft is not shown to the board and does not influence patient care.

Orthopedics and sarcoma board

Consecutive cases presented at the orthopedics and sarcoma tumor board of TUM University Hospital rechts der Isar.

A locally operated, source-grounded large language model system drafts a tumor board recommendation with justification and verified citations for each case before the board meeting. The draft is not shown to the board and does not influence patient care.

Head and neck board

Consecutive cases presented at the head and neck tumor board of TUM University Hospital rechts der Isar.

A locally operated, source-grounded large language model system drafts a tumor board recommendation with justification and verified citations for each case before the board meeting. The draft is not shown to the board and does not influence patient care.

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
Proportion of Cases With Category Match Between System Draft and Board Decision
時間枠:Each case is assessed once, after the board report of its session is available; analysis after accrual is complete (anticipated January 2027)

Proportion of discussed cases in which the system draft and the board consensus decision fall into the same of seven decision categories (systemic therapy, surgery, radiotherapy, watch and wait, further diagnostics, best supportive care, referral to the molecular tumor board). Overlap counts where either side carries a second category. Assessed by a pre-specified LLM judge (GLM-5.3-Flash, prompt V3, majority of three votes), pooled over the three boards. Tested for non-inferiority against the rate of the same system and judge on the retrospective reference sample of the same boards, margin 7.5 percentage points, one-sided alpha 0.025.

Each case is assessed once, after the board report of its session is available; analysis after accrual is complete (anticipated January 2027)

二次結果の測定

結果測定
メジャーの説明
時間枠
Proportion of Cases Rated Concordant or Partially Concordant
時間枠:After the board report of each session is available; analysis after accrual is complete (anticipated January 2027)

Category match with a substance score of at least 1 of 3 (concordant requires at least 2 of 3 on a primary-category match, 3 of 3 where only second categories overlap). Non-inferiority versus retrospective reference, margin 7.5 percentage points, tested after the primary outcome (hierarchical testing).

After the board report of each session is available; analysis after accrual is complete (anticipated January 2027)
Proportion of Cases Rated Concordant, Partially Concordant and Not Concordant
時間枠:After the board report of each session is available; analysis after accrual is complete (anticipated January 2027)

Proportion of cases rated concordant, partially concordant and not concordant by the LLM judge.

After the board report of each session is available; analysis after accrual is complete (anticipated January 2027)
Proportion of Cases With Category Match per Board
時間枠:After the board report of each session is available; analysis after accrual is complete (anticipated January 2027)

Category match and concordant or partially concordant per board, with 95% confidence intervals, descriptive.

After the board report of each session is available; analysis after accrual is complete (anticipated January 2027)
Proportion of Discordant Cases by Direction of Disagreement
時間枠:After the board report of each session is available; analysis after accrual is complete (anticipated January 2027)

Among cases without category match, proportion in which the draft is more conservative, more aggressive or a different modality than the board decision.

After the board report of each session is available; analysis after accrual is complete (anticipated January 2027)
Proportion of Cases With Category Match as Rated by Clinicians
時間枠:Within 3 months after the board session

In 30 to 50 cases per board, agreement between draft and board decision as rated by physicians who did not attend the session, and agreement between judge and clinicians on category match (percentage agreement, Cohen's kappa).

Within 3 months after the board session
Proportion of Citations Verified Word for Word Against Their Source
時間枠:At draft generation, before the board session

Proportion of citations per case that verify word for word against their source, and proportion of cases in which every citation verifies.

At draft generation, before the board session
Proportion of Cases With a Draft Completed Before the Board Session
時間枠:At draft generation, before the board session

Proportion of cases with a draft completed before the session starts, and generation time per case in minutes.

At draft generation, before the board session

その他の成果指標

結果測定
メジャーの説明
時間枠
Number of Clinical Trials Shortlisted per Case by the Trial-Matching Engine
時間枠:At draft generation, before the board session

Number of trials shortlisted per case, proportion recruiting on the board date, rule-based error rate (wrong disease, line of therapy, setting or status) and clinician screening verdicts where given.

At draft generation, before the board session

協力者と研究者

ここでは、この調査に関係する人々や組織を見つけることができます。

スポンサー

捜査官

  • 主任研究者:Lisa Adams, MD、Technical University of Munich
  • 主任研究者:Krischan Braitsch, MD、Technical University of Munich

研究記録日

これらの日付は、ClinicalTrials.gov への研究記録と要約結果の提出の進捗状況を追跡します。研究記録と報告された結果は、国立医学図書館 (NLM) によって審査され、公開 Web サイトに掲載される前に、特定の品質管理基準を満たしていることが確認されます。

主要日程の研究

研究開始 (推定)

2026年11月9日

一次修了 (推定)

2027年1月31日

研究の完了 (推定)

2027年4月30日

試験登録日

最初に提出

2026年11月9日

QC基準を満たした最初の提出物

2026年11月9日

最初の投稿 (実際)

2026年11月16日

学習記録の更新

投稿された最後の更新 (実際)

2026年11月16日

QC基準を満たした最後の更新が送信されました

2026年11月9日

最終確認日

2026年11月1日

詳しくは

本研究に関する用語

その他の研究ID番号

  • ONKOBOT-SHADOW-01
  • KK6050201BM5 (その他の助成金/資金番号:Central Innovation Programme for SMEs (ZIM), German Federal Ministry for Economic Affairs)

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?

いいえ

IPD プランの説明

Individual participant data are routine clinical records that may not leave the hospital under the ethics approval and data protection concept. Aggregate results and the analysis code will be published.

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

いいえ

米国FDA規制機器製品の研究

いいえ

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。