Agreement of ASA Physical Status Classification Among Anesthesia Residents, a Large Language Model, and Senior Anesthesiologist Consensus (ASA-AGREE)
Comparison of ASA Physical Status Classification by Anesthesia Residents, an Artificial Intelligence Large Language Model, and a Two-Anesthesiologist Expert Consensus: A Prospective Three-Rater Validation Study
The ASA Physical Status Classification is widely used in preoperative risk assessment, but agreement between raters is limited, particularly between ASA II and III. This prospective observational study compares the ASA class assigned to the same patients by three types of raters: anesthesia residents working in the preoperative anesthesia clinic, a large language model (LLM), and a consensus of two senior anesthesiologists, which is the reference standard.
Residents complete the standard preoperative evaluation form and assign an ASA class as part of routine care. A copy of the completed form, with patient identifiers and the resident's ASA class removed, is then given independently to the LLM and to the two senior anesthesiologists. No intervention is applied to participants and routine care is not changed.
The primary aim is to measure how closely the residents and the LLM agree with the expert consensus (weighted kappa). Secondary aims are agreement among all three raters, agreement between the two senior anesthesiologists, the direction of misclassification, and how often the LLM matches the consensus when the resident does not.
調査の概要
詳細な説明
Background The ASA Physical Status Classification is the most widely used tool for describing a patient's preoperative health status. It depends on clinical judgment, and agreement between raters is limited, especially between ASA II and III. In teaching hospitals many preoperative evaluations are done by residents. Large language models can read a structured clinical record and return a classification, and their agreement with anesthesiologists needs to be examined on real preoperative clinic data.
Design and setting Prospective, single-center, observational validation study with three rater groups, conducted in the preoperative anesthesia outpatient clinic of Kayseri City Hospital. Eligible adult patients scheduled for elective surgery are enrolled consecutively after written informed consent.
Procedures The resident takes the history, examines the patient, reviews laboratory results and consultation reports, completes the standard preoperative evaluation form and assigns an ASA class, as in routine care. The completed form is copied digitally. The resident's ASA class is stored in a locked field and is not shown to any other rater until all ratings are complete.
Two further copies are prepared in which patient identifiers (name, archive number, date) and the resident's ASA class, including any ASA checkbox and the free-text assessment, are masked. One copy goes to the LLM and the other to the two senior anesthesiologists.
The LLM is Claude Opus 5.5 (Anthropic), used through the Claude web interface (claude.ai), as a clinician would use it in practice. A dedicated study account is used, with memory and personalization turned off and no custom instructions. Each masked form image is entered in a new chat with the same fixed prompt. The model is queried once per patient, its response is saved verbatim with a screenshot, and the ASA class is entered into the database independently by two investigators. No patient identifiers are sent to the model.
The two senior anesthesiologists (more than 10 years of experience, not involved in the patient's perioperative care) classify each masked form independently. When they disagree, they discuss the case and record a consensus class and the reason. The consensus class is the reference standard. Agreement between the two experts before consensus is reported as a separate result.
Sample size Target enrollment is 330 patients, with 300 analyzable forms expected after allowing for 10% incomplete or excluded forms. The sample size was determined by Monte Carlo simulation for the primary comparison, the difference between two dependent quadratic weighted kappas (LLM vs consensus and resident vs consensus). Assuming a resident-consensus kappa of 0.45 and a clinically relevant difference of 0.15, 300 analyzable patients give 83% power at a two-sided alpha of 0.05.
Statistical analysis Agreement of the resident and of the LLM with the expert consensus is measured with quadratic weighted Cohen's kappa. The primary comparison is the difference between the two kappas, tested with a paired bootstrap 95% confidence interval (2000 resamples of patients). Agreement among all three raters is measured with Fleiss' kappa. Misclassification is described by direction (under- or over-classification) using confusion matrices. The ASA II-III boundary, resident training year and the inter-expert subgroup are pre-specified subgroup analyses. Predictors of resident misclassification are examined with logistic regression. All tests are two-sided at alpha 0.05; secondary comparisons are reported with 95% confidence intervals and treated as exploratory.
研究の種類
入学 (推定)
連絡先と場所
研究連絡先
- 名前:Çiğdem Ünal Kantekin, MD
- 電話番号:+90 352 315 77 00
- メール:drcgdm@gmail.com
研究場所
-
-
-
Kayseri、トルコ(Türkiye)、38080
- Kayseri City Hospital, Department of Anesthesiology and Reanimation
-
コンタクト:
- Çiğdem Ünal Kantekin, MD
- 電話番号:+90 352 315 77 00
- メール:drcgdm@gmail.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
サンプリング方法
調査対象母集団
Adult patients scheduled for elective surgery who attend the preoperative anesthesia outpatient clinic of Kayseri City Hospital and whose standard preoperative evaluation form is completed by a resident.
説明
Inclusion Criteria:
- Age 18 years or older
- Scheduled for elective surgery and evaluated at the preoperative anesthesia clinic
- Standard preoperative evaluation form completed by a resident, with required laboratory results and consultation reports entered
- Written informed consent
Exclusion Criteria:
- Emergency or semi-urgent surgery
- Incomplete preoperative evaluation form
- Unable to give informed consent (for example, because of a psychiatric disorder)
- Refusal to participate
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:COHORT
- 時間の展望:PROSPECTIVE
コホートと介入
グループ/コホート |
|---|
|
Preoperative anesthesia clinic patients
Adult patients scheduled for elective surgery who are evaluated at the preoperative anesthesia outpatient clinic. The anonymized preoperative form of each patient is rated for ASA class by a resident, an LLM and two senior anesthesiologists. |
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 |
|---|---|---|
|
Agreement between resident and expert-consensus ASA classification (quadratic weighted kappa)
時間枠:Once per patient, from the preoperative clinic evaluation until completion of the blinded ratings (up to 4 weeks)
|
Quadratic weighted Cohen's kappa between the ASA physical status class (I to V) assigned by the anesthesia resident and the consensus class assigned by two senior anesthesiologists who were blinded to the resident's class. The consensus class is the reference standard. A value of 1 indicates perfect agreement, 0 indicates chance agreement, and negative values indicate agreement below chance. |
Once per patient, from the preoperative clinic evaluation until completion of the blinded ratings (up to 4 weeks)
|
|
Agreement between large language model and expert-consensus ASA classification (quadratic weighted kappa)
時間枠:Once per patient, from the preoperative clinic evaluation until completion of the blinded ratings (up to 4 weeks)
|
Quadratic weighted Cohen's kappa between the ASA class assigned by the large language model from the anonymized preoperative form and the consensus class of two senior anesthesiologists. The model does not see the resident's class or any patient identifier. |
Once per patient, from the preoperative clinic evaluation until completion of the blinded ratings (up to 4 weeks)
|
二次結果の測定
結果測定 |
メジャーの説明 |
時間枠 |
|---|---|---|
|
Three-rater agreement of ASA classification (Fleiss' kappa)
時間枠:From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
Fleiss' kappa across the resident class, the large language model class and the expert consensus class for each patient. |
From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
|
Agreement between the two senior anesthesiologists before consensus (quadratic weighted kappa)
時間枠:From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
Quadratic weighted kappa between the independent ASA classes of senior anesthesiologist A and B, calculated before any discussion between them. |
From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
|
Agreement between resident and large language model ASA classification (quadratic weighted kappa)
時間枠:From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
Quadratic weighted kappa between the ASA class assigned by the resident and the class assigned by the large language model. |
From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
|
Direction of misclassification relative to expert consensus
時間枠:From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
Proportion of patients classified lower (under-classification) or higher (over-classification) than the expert consensus, reported separately for residents and for the large language model. The ASA II-III boundary is analyzed as a pre-specified subgroup. |
From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
|
Proportion of resident misclassifications in which the large language model matches expert consensus
時間枠:From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
Among patients whose resident ASA class differs from the expert consensus, the proportion for whom the large language model class equals the consensus class. This is an exploratory measure of decision-support potential. |
From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
|
Factors associated with resident misclassification
時間枠:From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
Odds ratios from logistic regression of resident misclassification (yes or no) on age, body mass index, number of ASA-relevant comorbidities, resident training year and borderline ASA II-III status. |
From the preoperative evaluation to completion of blinded ratings (up to 4 weeks per patient); assessed at the end of data collection (up to 8 months).
|
協力者と研究者
スポンサー
捜査官
- 主任研究者:Çiğdem Ünal Kantekin, MD、Kayseri City Hospital
研究記録日
主要日程の研究
研究開始 (推定)
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
その他の研究ID番号
- ASA-AGREE-2026-01
- 1208 (その他の識別子:Kayseri City Hospital Non-Interventional Clinical Research Ethics Committee)
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
Individual participant data will not be shared outside the study team. The consent form covers coded, confidential use of the data for this study only. Aggregate results 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
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。