Surface electromyography (sEMG) is a non-invasive technique that records the electrical signals generated by muscles through electrodes placed on the skin. sEMG is the state-of-the-art method used to control active upper limb prostheses because of the association between its amplitude and the neural drive sent from the spinal cord to muscles. However, accurately estimating the kinematics of a freely moving human hand using sEMG from extrinsic hand muscles remains a challenge. Deep learning has been recently successfully applied to this problem by mapping raw sEMG signals into kinematics. Nonetheless, the optimal number of EMG signals and the type of pre-processing that would maximize performance have not been investigated yet. Here, we analyze the impact of these factors on the accuracy in kinematics estimates. For this purpose, we processed monopolar sEMG signals that were originally recorded from 320 electrodes over the forearm muscles of 13 subjects. We used a previously published deep learning method that can map the kinematics of the human hand with real-time resolution. While state-of-the-art myocontrol algorithms essentially use the temporal envelope of the EMG signal as the only EMG feature, we show that our approach requires the full bandwidth of the signal in the temporal domain for accurate estimates. Spatial filtering had a smaller impact and low-order spatial filters may be suitable. Moreover, reducing the number of channels by ablation resulted in large performance losses. The highest accuracy was reached with the highest number of available sensors (320). Importantly and unexpected, our results also showed that increasing the number of channels above those used in this study may further enhance the accuracy in predicting the kinematics of the human hand. We conclude that full bandwidth high-density EMG systems of hundreds of electrodes are needed for accurate kinematic estimates of the human hand.