The standard 12-lead electrocardiogram (ECG) is widely used by cardiologists in diagnosing cardiac abnormalities. However, manual interpretation of ECG signals can be time consuming and dependent on the skills of the clinicians. In this work, we present an approach for detection of cardiac abnormalities through automatic analysis of 12-lead ECGs. The proposed technique is trained and validated on a dataset of 6,877 recordings with eight cardiac abnormalities and one normal sinus rhythm. Our approach applies median filtering to raw ECG signals and uses it as a direct input to two CNN-BiLSTM networks. The first network is trained for multiclass classification and the second network is trained for binary classification in a one-against-all strategy. A weighted classification layer is used to compensate the high imbalance in the class sizes. The probability outputs of these two networks are concatenated and used to train a softmax layer for making the final prediction. While the ECG signals in the dataset is up to 60 seconds long, our approach utilizes only the first 15 seconds of the signals as it was seen to produce comparable performance with lower computational costs. Our team (AIHI) achieved a F2 score of 0.826 and G2 score of 0.610 (geometric mean of 0.710) in four-fold cross-validation and a F2 score of 0.813 and G2 score of 0.590 (geometric mean of 0.693) on the hidden test set. Our current method learns temporal information from the raw ECG signals. In future, we plan to complement it with frequency domain characteristics to potentially achieve an even better classification performance.