标题： 将机器学习可解释性转化为重症监护病房死亡预测的临床见解 显示英文标题

标题： Translating Machine Learning Interpretability into Clinical Insights for ICU Mortality Prediction

摘要： 当前的研究努力主要集中在最多使用一种可解释的方法来阐明机器学习（ML）模型性能。 然而，将这些可解释性技术转化为临床医生的可操作见解仍存在重大障碍，特别是由于临床环境中的变异性以及Rashomon效应等复杂性。 在本研究中，我们开发并严格评估了两个ML模型及其解释机制，利用来自美国208家医院的131,051例重症监护病房（ICU）入院数据，数据来源为eICU协作研究数据库。 我们检查了两个数据集：一个包含填补缺失值的数据集（130,810名患者，5.58%的ICU死亡率）和一个排除缺失数据患者的另一个数据集（5,661名患者，23.65%的ICU死亡率）。 随机森林（RF）模型在第一个数据集上的AUROC为0.912，在第二个数据集上的AUROC为0.839，而XGBoost模型在第一个数据集上的AUROC为0.924，在第二个数据集上的AUROC为0.834。 在不同数据集、交叉验证折、模型和解释机制中，一致识别出的ICU死亡率预测因素包括乳酸水平、动脉pH值、体温等。 通过与常规收集的临床变量对齐，本研究旨在提高ML模型的可解释性以用于临床，促进临床医生更深入的理解和采用，并最终有助于改善患者结果。 显示英文摘要

摘要： Current research efforts largely focus on employing at most one interpretable method to elucidate machine learning (ML) model performance. However, significant barriers remain in translating these interpretability techniques into actionable insights for clinicians, notably due to complexities such as variability across clinical settings and the Rashomon effect. In this study, we developed and rigorously evaluated two ML models along with interpretation mechanisms, utilizing data from 131,051 ICU admissions across 208 hospitals in the United States, sourced from the eICU Collaborative Research Database. We examined two datasets: one with imputed missing values (130,810 patients, 5.58% ICU mortality) and another excluding patients with missing data (5,661 patients, 23.65% ICU mortality). The random forest (RF) model demonstrated an AUROC of 0.912 with the first dataset and 0.839 with the second dataset, while the XGBoost model achieved an AUROC of 0.924 with the first dataset and 0.834 with the second dataset. Consistently identified predictors of ICU mortality across datasets, cross-validation folds, models, and explanation mechanisms included lactate levels, arterial pH, body temperature, and others. By aligning with routinely collected clinical variables, this study aims to enhance ML model interpretability for clinical use, promote greater understanding and adoption among clinicians, and ultimately contribute to improved patient outcomes.