The aim of this study was to develop and compare techniques to increase the prediction accuracy of patient mortality and organ dysfunction in the Intensive Care Units (hereinafter ICU) of hospitals. Patient mortality was estimated with two models of artificial neural network (ANN)-backpropagation (BP) and simplified acute physiology score (SAPS). Organ dysfunction was predicted by coupled ANN self-organizing map (SOM) and logistic organ dysfunction score (LODS) method on the basis of patient conditions. Input dataset consisted of 36 features recorded for 4,000 patients in the ICU. An integrated response surface methodology (RSM) and genetic algorithm (GA) was developed to achieve the best topology of the ANN-BP model. Although mortality prediction of the best ANN-BP (MSE = 0.0036, AUC = 0.83, R² = 0.81) was more accurate than that of the SAPS score model (MSE = 0.0056, AUC = 0.82, R² = 0.78), the execution time of the former (=45 min) was longer than that of the latter (=20 min). Therefore, the principal component analysis (PCA) was used to reduce the input feature dimensions, which, in turn, reduced the execution time up to 50%. Data reduction also helped to increase the network accuracy up to 90%. The likelihood of organ dysfunction determined by coupled ANN and scoring method technique can be much more efficient than the LODS model alone because the SOM could successfully classify the patients in 64 classes. The primary patient classification plays a major role in increasing the efficiency of an estimator.

中文翻译：

---

集成人工神经网络和评分系统，以提高患者死亡率和器官功能障碍的预测准确性。

这项研究的目的是开发和比较技术，以提高医院的重症监护病房（以下简称ICU）患者死亡率和器官功能障碍的预测准确性。使用两种人工神经网络（ANN）反向传播（BP）模型和简化的急性生理学评分（SAPS）模型估计患者的死亡率。通过结合ANN自组织图（SOM）和后勤器官功能障碍评分（LODS）方法，根据患者情况预测器官功能障碍。输入数据集包括为ICU中的4,000名患者记录的36个特征。为了实现ANN-BP模型的最佳拓扑，开发了一种综合的响应面方法（RSM）和遗传算法（GA）。尽管最佳ANN-BP的死亡率预测（MSE = 0.0036，AUC = 0.83，R ² = 0.81）比SAPS评分模型（MSE = 0.0056，AUC = 0.82，R ²  = 0.78）更准确，前者（= 45分钟）的执行时间比后者（= 20分钟） ）。因此，使用主成分分析（PCA）来减少输入要素的尺寸，从而将执行时间减少多达50％。数据减少还有助于将网络准确性提高到90％。由于SOM可以成功地将患者分类为64个类别，因此通过ANN和评分方法耦合确定的器官功能障碍的可能性比单独使用LODS模型更为有效。主要患者分类在提高估计器效率方面起着重要作用。