A brain-computer interface (BCI) system based on event-related potentials (ERP) can be used to measure human intent for improved motor function rehabilitation applications. Somatosensory ERP is a paradigm especially suitable for patients with visual or hearing impairment, yet few algorithms were dedicated to it. In this study we conducted ERP experiments on healthy subjects with somatosensory (facilitated by electric stimuli), audio, and visual sessions. A convolutional neural network (CNN) classifier was specifically tuned to optimize classification accuracy of somatosensory ERP. The network includes temporal-spatial processing layers followed by MobileNet-styled module with Squeeze-Excitation. The model was trained as subject-specific classifier. The effect of using multiple ERP trials for classification and reduced number of EEG channels was also evaluated. For the standard single-Trial 64-channel configuration, the proposed model achieved 81.4% accuracy for electric stimuli data, higher than the baseline models by 1.4 to 3.3%. Switching from 64 to 8-channel configuration brought only 3.2% decreased accuracy while further reducing to 4-channel did not make significant changes. Increasing the number of ERP trials used for classification improves the accuracy, which plateaued around 90% at 5-Trial. In terms of implementation, the proposed model brought the classification accuracy improvement without significantly increasing the model size or inference time tested on mobile platforms. Ablation study showed that the presented architecture was the optimal among evaluated variations. For future development, the proposed model and somatosensory ERP paradigm will be assessed in terms of user intent recognition on real patients, and the model architecture will be continuously optimized for designated BCI task and dataset.