Cardiac arrest has a lower survival rate.It is important to predict cardiac arrest as early as possible in order to act and prevent it from occurring.There are numerous machine learning and deep learning algorithms to predict cardiac arrest but due to low sensitivity and a high false alarm rate,they have not been successfully implemented.The goal of this research is to lower the rate of false alarm and improve the accuracy,sensitivity and specificity of early warning systems in hospital for a sepsis patient who are at risk of cardiac arrest.We offer a deep learning-based early prediction model.The suggested system is based on the MIMIC-III database,which generates two types of data:timeseries data and a mix of time series and baseline data.The system model is made up of two upgraded Bidirectional LSTM（Long Short-Term Memory）and a hybrid model called MLP（Multilayer Perceptron）.MLP handles baseline variables such as age,sex,and primary complaints,whereas bidirectional LSTM handles time series vital signs data in both forward and backward directions.Missing values,inconsistencies in time series,and an imbalanced data collection are also addressed which improves the accuracy of this research.The Model can predict cardiac arrest up to six hours before it occurs.The proposed system improves upon the conventional machine learning and neural network algorithms,acuracy,sensitivity,specificity,and Area Under Curve（AUC）of the traditional model（state-of-the-art）model are 85.3%,87.4%,and 84.2%,0.83,respectively,while accuracy,sensitivity,specificity,and AUC of the suggested solution are 92.5%,94.2%,93.1%,0.92,respectively.