Mortality prediction in intensive care units with the Super ICU Learner Algorithm (SICULA): a population-based study.
Mortality prediction in intensive care units with the Super ICU Learner Algorithm (SICULA): a population-based study.
复制标题
DOI：
10.1016/s2213-2600(14)70239-5
复制
发表时间：
2015-01
影响因子：
76.2
通讯作者：
van der Laan, Mark J.
中科院分区：
文献类型：
--
作者：
Pirracchio, Romain;Petersen, Maya L.;Carone, Marco;Rigon, Matthieu Resche;Chevret, Sylvie;van der Laan, Mark J.
改善重症监护病房 (ICU) 患者的死亡率预测仍然是一个重要的挑战。已经提出了许多严重程度评分，但验证研究得出的结论是它们没有得到充分校准。有许多灵活的算法可用，但无论上下文如何，这些算法都没有单独优于所有其他算法。相比之下，超级学习器（SL）是一种集成机器学习技术，利用多种学习算法来获得更好的预测性能，已被证明至少与其库中的最佳成员一样好。它可能提供一个理想的机会来构建具有改进的性能概况的新颖的严重性评分。本研究的目的是使用超级学习器的实现为 ICU 患者提供一种新的死亡率预测算法，并根据 SAPS II、APACHE II 和 SOFA 评分评估其相对于预测的性能。我们使用重症监护 II (MIMIC-II) 数据库 (v26) 中的多参数智能监测 (MIMIC-II) 数据库，包括 2001 年至 2008 年入住波士顿 Beth Israel Deaconess 医疗中心 ICU 的所有患者。对基于 SAPS II、APACHE II、SOFA 和我们基于超级学习的提案的预测医院死亡率的校准、区分和风险分类进行了评估。使用交叉验证来计算绩效指标，以避免做出有偏见的评估。然后，我们提出的评分在 2013 年 9 月至 2014 年 6 月期间入住法国巴黎欧洲乔治蓬皮杜医院 ICU 的 200 名随机选择患者的数据集上进行了外部验证。主要结局是医院死亡率。解释变量与 SAPS II 评分中包含的变量相同。纳入 24,508 名患者，中位 SAPS II 38（IQR：27-51），中位 SOFA 5（IQR：2-8）。共有3,002/24,508名患者（12.2%）在医院死亡。我们基于超级学习者的提案的两个版本得出的平均预测死亡概率分别为 0.12（IQR：0.02-0.16）和 0.13（IQR：0.01-0.19），而 SOFA 和 SAPS II 分数的相应值分别为 0.12（IQR：0.05-0.15）和 0.30（IQR：0.01-0.19）。 0.08–0.48）。 SAPS II 和 SOFA 的受试者工作特征曲线 (AUROC) 下的交叉验证面积分别为 0.78 (95% CI: 0.77–0.78) 和 0.71 (95% CI: 0.71–0.72)。当解释变量按照 SAPS II 进行分类时，我们的建议达到了 0.85（95%CI：0.84-0.85）的 AUROC；当包含相同的解释变量而不进行任何转换时，我们的建议达到了 0.88（95%CI：0.87-0.89）。此外，它还表现出比以前的评分系统更好的校准特性。在外部验证数据集上，AUROC 为 0.94（95%CI：0.90-0.98），校准特性良好。与传统的严重程度评分相比，我们基于超级学习者的提案在预测 ICU 患者的医院死亡率方面提供了改进的性能。用户友好的实施方式可在线获取，对于寻求验证我们分数的临床医生来说应该是有用的。富布赖特基金会、公共援助 – 巴黎医院 (RP)；多丽丝·杜克 (Doris Duke) 临床科学家发展奖 (MP) 和 NIH 拨款 # 2R01AI074345-06A1(MvdL)。
Improved mortality prediction for patients in intensive care units (ICU) remains an important challenge. Many severity scores have been proposed but validation studies have concluded that they are not adequately calibrated. Many flexible algorithms are available, yet none of these individually outperform all others regardless of context. In contrast, the Super Learner (SL), an ensemble machine learning technique that leverages on multiple learning algorithms to obtain better prediction performance, has been shown to perform at least as well as the optimal member of its library. It might provide an ideal opportunity to construct a novel severity score with an improved performance profile. The aim of the present study was to provide a new mortality prediction algorithm for ICU patients using an implementation of the Super Learner, and to assess its performance relative to prediction based on the SAPS II, APACHE II and SOFA scores. We used the Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC-II) database (v26) including all patients admitted to an ICU at Boston’s Beth Israel Deaconess Medical Center from 2001 to 2008. The calibration, discrimination and risk classification of predicted hospital mortality based on SAPS II, on APACHE II, on SOFA and on our Super Learned-based proposal were evaluated. Performance measures were calculated using cross-validation to avoid making biased assessments. Our proposed score was then externally validated on a dataset of 200 randomly selected patients admitted at the ICU of Hôpital Européen Georges-Pompidou in Paris, France between September 2013 and June 2014. The primary outcome was hospital mortality. The explanatory variables were the same as those included in the SAPS II score. 24,508 patients were included, with median SAPS II 38 (IQR: 27–51), median SOFA 5 (IQR: 2–8). A total of 3,002/24,508(12.2%) patients died in the hospital. The two versions of our Super Learner-based proposal yielded average predicted probabilities of death of 0.12 (IQR: 0.02–0.16) and 0.13 (IQR: 0.01–0.19), whereas the corresponding values for the SOFA and SAPS II scores were, respectively, 0.12 (IQR: 0.05–0.15) and 0.30 (IQR: 0.08–0.48). The cross-validated area under the receiver operating characteristics curve (AUROC) for SAPS II and SOFA were 0.78(95%CI: 0.77–0.78) and 0.71 (95%CI: 0.71–0.72), respectively. Our proposal reached an AUROC of 0.85 (95%CI: 0.84–0.85) when the explanatory variables were categorized as in SAPS II, and of 0.88 (95%CI: 0.87–0.89) when the same explanatory variables were included without any transformation. In addition, it exhibited better calibration properties than previous score systems. On the external validation dataset, the AUROC was 0.94 (95%CI: 0.90–0.98) and calibration properties were good. As compared to conventional severity scores, our Super Learner-based proposal offers improved performance for predicting hospital mortality in ICU patients. A user-friendly implementation is available online and should prove useful to clinicians seeking to validate our score. Fulbright Foundation, Assistance Publique – Hôpitaux de Paris (RP); Doris Duke Clinical Scientist Development Award (MP) and NIH Grant # 2R01AI074345-06A1(MvdL).