Cardiac arrest is a common issue in Intensive Care Units (ICU) with low survival rate. Deep learning algorithms have been used to predict cardiac arrest unsuccessfully due to low sensitivity and high false alarm rate. The aims of this research are to decrease the false alarm rate and increase sensitivity and specificity. The proposed system is based on Medical Information Mart for Intensive Care (MIMIC-III) database where two sets of data are created. These two datasets are time-series data and combination of time series and baseline data. Time series dataset is divided into six-time groups. The system model consists of a hybrid model: Multilayer Perceptron (MLP) and enhanced Bidirectional Long Short-Term Memory (LSTM). MLP processes baselines feature like age, sex, chief complaints whereas the bidirectional LSTM is used to handle time series vital signs data from forward and backward direction so that it considers both present and future inputs. The model predicts cardiac arrest up to six hours earlier before the incidence. We achieved better performance for combined dataset where the prediction time window is 1 h. Accuracy, sensitivity, specificity, and Area Under Curve (AUC) equal to 85.7%, 87.7%,84.9%, and 0.86 respectively for the state of art, for proposed solution are 92.6%, 94.3%, 93.6% and 0.94 respectively. The proposed system is reducing the false alarm rate and increasing accuracy, sensitivity, specificity, and the area under curve for the prediction of cardiac arrest using enhanced Bidirectional LSTM model. The problem of missing values, irregularities of time series, and imbalance data set is solved too.

心脏骤停是重症监护病房 (ICU) 中生存率低的常见问题。由于低灵敏度和高误报率，深度学习算法已被用于预测心脏骤停失败。本研究的目的是降低误报率并提高灵敏度和特异性。提议的系统基于重症监护医疗信息集市 (MIMIC-III) 数据库，其中创建了两组数据。这两个数据集是时间序列数据以及时间序列和基线数据的组合。时间序列数据集分为六个时间组。该系统模型由一个混合模型组成：多层感知器 (MLP) 和增强型双向长短期记忆 (LSTM)。MLP 处理基线特征，如年龄、性别、主要抱怨而双向 LSTM 用于处理来自前向和后向的时间序列生命体征数据，以便它考虑当前和未来的输入。该模型可在发病前最多提前 6 小时预测心脏骤停。我们在预测时间窗口为 1 小时的组合数据集上取得了更好的性能。对于现有技术，准确度、灵敏度、特异性和曲线下面积 (AUC) 分别等于 85.7%、87.7%、84.9% 和 0.86，建议的解决方案分别为 92.6%、94.3%、93.6% 和 0.94。所提出的系统正在降低误报率并提高准确性、灵敏度、特异性和使用增强型双向 LSTM 模型预测心脏骤停的曲线下面积。缺失值问题，时间序列不规则，