Development and Validation of a Cirrhosis-specific Surgical Risk Calculator (C-SuRC)

背景：肝硬化患者的围手术期死亡率是非肝硬化患者的2-4倍 肝硬化是由于与肝硬化相关的因素，如门静脉高压和肝脏合成功能受损。 目前还没有模型可以准确估计围手术期死亡率和发病率， 肝硬化我们的首要目标是开发和验证肝硬化特定手术风险计算器（C- SuRC），准确估计肝硬化患者的围手术期死亡率和并发症。 意义/影响：C-SuRC将改善肝硬化患者外科手术的选择， 为低死亡率患者提供择期手术，防止高死亡率患者接受手术， 确定可改变的风险因素，这些因素可以在手术前优化，以改善结果。 创新： · C-SuRC将是第一个专门为肝硬化患者设计的手术风险计算器， 合并了所有三个主要类别的预测因子，这些预测因子对以下患者的手术死亡率有贡献： 肝硬化，即与肝硬化相关的、与手术相关的和与合并症相关的预测因子。 · C-SuRC将使用我们通过合并VASQIP和CDW开发的独特数据集进行开发 数据这是一个全国性的代表性的VA数据集，患者接受外科手术 前瞻性收集基线特征和手术结果。 ·我们将开发和比较传统的逻辑回归模型以及最先进的， 梯度增强（XGBoost）机器学习算法。 ·我们将使用一种新的方法来解释机器学习算法（SHAP）的预测， 分配每个风险因素对模型预测的死亡率的贡献。这具有深刻的 对医疗预测分析中“可解释的人工智能”的影响。SHAP值可以用来“解释” 预测并识别可以在手术前改善的潜在可修改因素。 ·我们将应用以用户为中心的设计，开发基于Web和基于应用程序的工具，执行C-SuRC。 具体目标： SA 1.开发并外部验证可准确估计术后30天的模型（C-SuRC） 肝硬化患者的死亡率和并发症， 共病相关和手术相关的预测因子。 SA 2.使用一种新的方法（SHapley加法解释或“SHAP”）来计算 通过我们的C-SuRC梯度增强的机器学习模型预测的死亡风险的每个风险因素 在个别患者中。 SA 3.结合用户的反馈，并应用以用户为中心的设计中的最佳实践来开发Web- 基于和基于应用程序的工具，执行C-SuRC并显示手术结果预测， 个体患者和每个关键风险因素对使用SHAP值预测的风险的贡献。 方法：我们将使用传统的逻辑回归模型和最先进的梯度增强机 C-SuRC开发的学习模型。我们将测试C-SuRC的鉴别力、校准和准确性， 外部验证并将其与现有手术风险计算器进行比较。我们将使用SHAP值来计算 每个风险因素对机器学习模型预测的死亡率的贡献。我们将合并 来自25名临床医生用户的反馈，以开发执行C-SuRC的基于网络和基于应用程序的工具。 下一步/实施：我们将寻求所有重要的VA利益相关者的支持，其中许多人已经 我已经批准了这一建议，并传播我们的调查结果和基于网络和应用程序的C-SuRC 在全国范围内，VA工具作为肝硬化患者术前评估的常规工具。

Background: Perioperative mortality is 2-4 times higher in patients with cirrhosis compared to patients without cirrhosis due to cirrhosis-related factors such as portal hypertension and impaired hepatic synthetic function. Currently no models exist that accurately estimate peri-operative mortality and morbidity in patients with cirrhosis. Our overarching aim is to develop and validate a Cirrhosis-specific Surgical Risk Calculator (C- SuRC) that accurately estimates perioperative mortality and complications in patients with cirrhosis. Significance/Impact: C-SuRC will improve the selection of patients with cirrhosis for surgical procedures, improve access to elective surgery for patients with low mortality, prevent surgeries in patients with high mortality and identify modifiable risk factors that could be optimized prior to surgery in order to improve outcomes. Innovation: • C-SuRC will be the first surgical risk calculator specifically designed for patients with cirrhosis that incorporates all three major classes of predictors that contribute to operative mortality in patients with cirrhosis, that is cirrhosis-related, surgery-related and comorbidity-related predictors. • C-SuRC will be developed using a unique, dataset that we developed by merging VASQIP and CDW data. This is a nationally-representative VA dataset of cirrhotic patients undergoing surgical procedures with prospectively collected baseline characteristics and surgical outcomes. • We will develop and compare both traditional logistic regression models as well as state-of-the-art, gradient-boosted (XGBoost) machine learning algorithms. • We will use a novel method for interpreting the predictions of machine learning algorithms (SHAP), which assigns the contribution of each risk factor to the mortality predicted by the model. This has profound implications for “interpretable AI” in medical predictive analytics. SHAP values can be used to “explain” a prediction and to identify potentially modifiable factors that can be improved prior to surgery. • We will apply user-centered design to develop web-based and app-based tools that execute C-SuRC. Specific Aims: SA1. Develop and externally validate a model (C-SuRC) that accurately estimates 30-day postoperative mortality and complications in patients with cirrhosis using routinely available cirrhosis-related, comorbidity-related and surgery-related predictors. SA2. Use a novel method (the SHapley Additive exPlanations or “SHAP”) to calculate the contribution of each risk factor to the mortality risk predicted by our C-SuRC gradient boosted, machine learning models in individual patients. SA3. Incorporate feedback from users and apply best practices in user-centered design to develop web- based and app-based tools that execute C-SuRC and display predictions of surgical outcomes in individual patients and the contribution of each key risk factor to the predicted risk using SHAP values. Methods: We will use conventional logistic regression models and state-of-the-art, gradient-boosted machine learning models for C-SuRC development. We will test the discrimination, calibration and accuracy of C-SuRC, externally validate it and compare it to existing surgical risk calculators. We will use SHAP values to calculate the contribution each risk factor to the mortality predicted by the machine learning models. We will incorporate feedback from 25 clinician-users to develop web-based and app-based tools that execute C-SuRC. Next Steps/Implementation: We will solicit support from all important VA stakeholders, many of whom have already endorsed this proposal, and disseminate our findings and the web-based and app-based C-SuRC tools in the VA nationally as a routine instrument in the pre-operative assessment of patients with cirrhosis.