About 50% of all cardiovascular-related deaths are the result of a sudden cardiac death (SCD) event, which is the abrupt loss of consciousness within less than one hour after the onset of symptoms. While the SCD is the ultimate event, it is triggered by other cardiac events characterized by alterations in the normal rhythm of the heart. This research explores a novel approach for analyzing the electrocardiogram (ECG) signals and machine learning algorithms to predict events such as ventricular fibrillation and generate alerts early enough to allow access to healthcare. This approach extracts the Shannon entropy as a time-variable feature from the ECG samples organized on multiple time-windows, enhancing SCD prediction in terms of lead time and accuracy respect to current results reported in literature. The experiments in this study were made with ECG records from the Physionet public ECG databases, taking the samples and annotations to train and evaluate several supervised learning algorithms, such as binary decision trees, convolutional neural networks (CNN), and support vector machines (SVM). Results show that support vector machines (SVM) have the best performance to predict the SCD event, with high performance metrics evaluated 15 min before the onset. Results show that calculating Shannon entropy from raw ECG signals has a better performance than other time-domain features extracted from the ECG, and from heart rate variability (HRV) features, allowing for prediction up to 15 min before the SCD event happens reaching 92.5% accuracy. The proposed model, trained with the SVM algorithm, demonstrates a competitive and improved classification accuracy and F1 score (95%), measured for 5 min before the onset to compare it with existing methods (92.31%). While most of the referenced papers report results in a time span of 5 min, this research contributes to the advancement of automatic SCD prediction techniques extending the time up to 15 min with a good prediction, detailing the method to replicate it with longer times. As the model works with one feature to represent the ECG signal, the computational requirements will be lower for potential applications in wearable monitoring devices.