Aims: Conventional Early Warning Scoring (EWS) systems can improve the quality of clinical care by providing heuristics for detection of abnormalities in high-risk patients. This study aims at outperforming the performance of common EWSs by devising a more advanced EWS system using Convolutional Bi-directional Long-Short Term Memory (CB-LSTM) neural networks learned on historical electronic medical records (EMR) of patients in intensive care units. Methods: First, the development data-set (DS) is enriched by adding all septic patients of extended training sets, ‘A’ and ‘B’, and 200 randomly selected non-septic ones from each set. Imputation of missing values in vital signs and laboratory values in EMR of each patient is carried out using the last known valid value of each parameter for each patient. Next, at each time-step, mean, standard deviation, and differences with the last 4 consecutive values of each parameter is calculated as features. Additionally, for all vital signs and glucose values, the difference between current value of parameter at each time-step and its expected value which is predicted using a Kalman model learned over the last 4 consecutive values is also calculated as feature. The missing Kalman features are filled with zero indicating no outlier. Moreover, Sepsis-related Organ Failure Assessment (SOFA) score, quick SOFA score and National Early Warning Score (NEWS) are calculated as features at each time-step. The final feature vector is constructed by excluding those parameters and their corresponding statistics, with less than 25% presence in the DS before the imputation step. A CB-LSTM neural networks with 2 convolutional layers followed with one bi-directional recurrent layer is trained over the DS. Results: The normalized utility score (NUS) of 0.22% is achieved in the unofficial phase. In addition, the overall NUS of 0.34% is achieved using 5-fold cross validation over the DS outperforming calculated SOFA scores and NEWS.