In this paper, the performance of a deep learning based multiple-input multiple-output (MIMO) non-orthogonal multiple access (NOMA) system is investigated for 5G radio communication networks. We consider independent and identically distributed (i.i.d.) Nakagami-m fading links to prove that when using MIMO with the NOMA system, the outage probability (OP) and end-to-end symbol error rate (SER) improve, even in the presence of imperfect channel state information (CSI) and successive interference cancellation (SIC) errors. Further more, the stacked long short-term memory (S-LSTM) algorithm is employed to improve the system’s performance, even under time-selective channel conditions and in the presence of terminal’s mobility. For vehicular NOMA networks, OP, SER, and ergodic sum rate have been formulated. Simulations show that an S-LSTM-based DL-NOMA receiver outperforms least square (LS) and minimum mean square error (MMSE) receivers. Furthermore, it has been discovered that the performance of the end-to-end system degrades with the growing amount of node mobility, or if CSI knowledge remains poor. Simulated curves are in close agreement with the analytical results.