Classifying limb movements using brain activity is an important task in Brain-computer Interfaces (BCI) that has been successfully used in multiple application domains, ranging from human-computer interaction to medical and biomedical applications. This paper proposes a novel solution for classification of left/right hand movement by exploiting a Long Short-Term Memory (LSTM) network with attention mechanism to learn from sequential data available in the electroencephalogram (EEG) signals. In this context, a wide range of time and frequency domain features are first extracted from the EEG signal and are then evaluated using a Random Forest (RF) to select the most important features. The selected features are arranged as a spatio-temporal sequence to feed the LSTM network, learning from the sequential data to perform the classification task. We conduct extensive experiments with the EEG motor movement/imagery database and show that our proposed solution achieves effective results outperforming baseline methods and the state-of-the-art in both intra-subject and cross-subject evaluation schemes. Moreover, we utilize the proposed framework to analyze the information as received by the sensors and monitor the activated regions of the brain by tracking EEG topography throughout the experiments.