AI Treatment Recommendations With Predicted Consequences for Sepsis Fluid and Vasopressor Decisions
Effect of Displaying Predicted Consequences of AI Treatment Recommendations on Clinicians' Fluid and Vasopressor Decisions in Sepsis: A Randomized Crossover Study
When artificial intelligence (AI) recommends a treatment, clinicians must decide whether to follow it. This study tests whether showing the predicted consequences of a recommendation, in addition to the recommendation itself, helps clinicians accept appropriate advice and reject inappropriate advice.
We developed a clinician-anchored counterfactual model that compares the on-duty clinician's intravenous fluid and vasopressor plan with alternative plans for patients with sepsis. It predicts the patient's course under each plan and recommends a change only when the predicted benefit holds across several plausible assumptions about unrecorded factors; otherwise it states that the evidence is insufficient to recommend a change.
About 72 physicians from two hospitals, stratified by seniority, will review 30 retrospective, de-identified ICU decision points from patients with sepsis. In stage 1, each physician independently specifies fluid and vasopressor doses without any AI information. In stage 2, the same cases are reviewed under three conditions assigned by counterbalanced sequences: the case with the on-duty clinician's plan and no AI (B0); B0 plus the AI recommendation (B1); and B1 plus the predicted consequences of the on-duty and recommended plans (B2). The AI recommendation is identical in B1 and B2. Decision quality is assessed against plans made independently by expert intensivists. No patients are enrolled, and participants' decisions do not affect patient care.
調査の概要
状態
詳細な説明
Cases. Thirty decision points were selected from retrospective, de-identified ICU records before the experiment. Sampling was stratified so that the set includes AI recommendations to change treatment and to keep the current plan, both appropriate and inappropriate according to the expert reference. Each patient contributes one decision point, and none of the cases were used to develop the AI system.
Expert reference. A panel of intensivists who did not take part in AI development or in the experiment independently specify a fluid and vasopressor plan for every case, without AI output and without knowledge of study conditions. A participant's plan is scored by the number of dose levels by which it lies outside the range of expert plans (0 if within the range), summed over the two drugs. AI recommendations are classified as appropriate or inappropriate by the same rule before participant data are analyzed.
Analysis. Outcomes are analyzed with mixed-effects models with physician and case as crossed random effects. Secondary outcomes are tested in a prespecified order after the primary outcome to control the overall type I error.
Related non-interventional components. A retrospective comparison of clinicians' and the AI's plans and a prospective shadow deployment, in which the system ran on live ICU data without influencing treatment, are reported alongside this experiment but are not part of the randomized comparison.
研究の種類
入学 (推定)
段階
- 適用できない
連絡先と場所
研究場所
-
-
Chongqing Municipality
-
Chongqing、Chongqing Municipality、中国、400016
- 募集
- The First Affiliated Hospital of Chongqing Medical University
-
コンタクト:
- Muchan Tao
- 電話番号:+86 18980918483
- メール:orchilipairs@gmail.com
-
-
Sichuan
-
Luzhou、Sichuan、中国、646000
- 募集
- The Affiliated Hospital of Southwest Medical University
-
コンタクト:
- Muchan Tao
- 電話番号:+86 18980918483
- メール:mooochann@163.com
-
-
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
説明
Inclusion Criteria:
- Licensed physicians in critical care, emergency medicine, anaesthesiology, or respiratory medicine.
- Currently treating ICU patients.
- Able to complete the online questionnaire.
Exclusion Criteria:
- Involvement in developing or evaluating the AI system, or in case selection
- Membership of the expert reference panel
- Participation in the pilot survey
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
- 主な目的:ヘルスサービス研究
- 割り当て:ランダム化
- 介入モデル:クロスオーバー割り当て
- マスキング:独身
武器と介入
参加者グループ / アーム |
介入・治療 |
|---|---|
|
実験的:Sequence A
Stage 1: all 30 cases answered without AI information. Stage 2 (questionnaire version A): cases 1-10 under B0, cases 11-20 under B1, cases 21-30 under B2. |
Participants review the case together with the on-duty clinician's actual intravenous fluid rate and vasopressor dose and submit a final plan. No AI information is shown. This condition controls for the effect of a second review. As B0, plus the AI output: either a recommended dose level for intravenous fluid and for vasopressor, shown with the corresponding dose range, or a statement that the evidence is insufficient to recommend a change. As B1, with identical AI output, plus the model's predicted mean arterial pressure, lactate and urine output over the next 4 hours under the on-duty plan and under the recommended plan, shown as point estimates. |
|
実験的:Sequence B
Stage 1: all 30 cases answered without AI information. Stage 2 (questionnaire version B): cases 1-10 under B1, cases 11-20 under B2, cases 21-30 under B0. |
Participants review the case together with the on-duty clinician's actual intravenous fluid rate and vasopressor dose and submit a final plan. No AI information is shown. This condition controls for the effect of a second review. As B0, plus the AI output: either a recommended dose level for intravenous fluid and for vasopressor, shown with the corresponding dose range, or a statement that the evidence is insufficient to recommend a change. As B1, with identical AI output, plus the model's predicted mean arterial pressure, lactate and urine output over the next 4 hours under the on-duty plan and under the recommended plan, shown as point estimates. |
|
実験的:Sequence C
Stage 1: all 30 cases answered without AI information. Stage 2 (questionnaire version C): cases 1-10 under B2, cases 11-20 under B0, cases 21-30 under B1. |
Participants review the case together with the on-duty clinician's actual intravenous fluid rate and vasopressor dose and submit a final plan. No AI information is shown. This condition controls for the effect of a second review. As B0, plus the AI output: either a recommended dose level for intravenous fluid and for vasopressor, shown with the corresponding dose range, or a statement that the evidence is insufficient to recommend a change. As B1, with identical AI output, plus the model's predicted mean arterial pressure, lactate and urine output over the next 4 hours under the on-duty plan and under the recommended plan, shown as point estimates. |
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 |
|---|---|---|
|
Distance of final treatment decisions from the range of expert plans, B2 versus B1
時間枠:At completion of the participant's stage-2 review (single online session, approximately 2 hours)
|
For each case, the number of dose levels by which the participant's final fluid and vasopressor plan lies outside the range of plans made independently by the expert panel (0 = within range), summed over the two drugs. Compared between B2 and B1 with a mixed-effects model with physician and case as crossed random effects, adjusted for stage-1 score, block position and sequence. |
At completion of the participant's stage-2 review (single online session, approximately 2 hours)
|
二次結果の測定
結果測定 |
メジャーの説明 |
時間枠 |
|---|---|---|
|
Discrimination between appropriate and inappropriate AI recommendations, B2 versus B1
時間枠:At completion of the participant's stage-2 review (single online session, approximately 2 hours)
|
Among cases in which the AI recommends changing treatment, acceptance is defined as a final plan equal to the AI-recommended dose level for the changed drug. The outcome is the interaction between condition (B2 vs B1) and recommendation appropriateness (classified by the expert-range rule before analysis) on acceptance, estimated with a mixed-effects probit model with physician and case as crossed random effects. |
At completion of the participant's stage-2 review (single online session, approximately 2 hours)
|
協力者と研究者
研究記録日
主要日程の研究
研究開始 (実際)
一次修了 (推定)
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
キーワード
追加の関連 MeSH 用語
その他の研究ID番号
- KY2026132
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?
IPD プランの説明
De-identified participant-level responses (treatment decisions, confidence, response times, condition assignments and participant characteristics without identifiers) will be shared. Patient case material derived from source databases will be shared only as permitted by their data use agreements.
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。