Abstract: In high-Earth-orbit and deep-space links， the long propagation distance and limited transmit power often drive navigation messages into an ultralow signal-to-noise ratio （SNR） regime at the receiver in which the reliability of low-density parity-check （LDPC） decoding degrades markedly and decoding becomes challenging. To address this issue， we propose an LDPC decoding approach that fuses navigation-domain priors with prediction by deep learning models. Building on the intrinsic temporal evolution of navigation parameters， the method operates within an extended min-sum （EMS） decoding framework and uses a long short-term memory （LSTM） network to predict orbit and clock-bias states. An evolution-consistency constraint is introduced to align the predicted trends with the physical dynamics of these parameters. The resulting priors are then used to re-weight and refine the initial variable-node log-likelihood ratios （LLR） to improve decoding robustness under low SNR. We validated the algorithm via simulation on the IGS RINEX 4.00 （2024） dataset using BeiDou satellites in medium Earth orbit （MEO） and inclined geosynchronous orbit （IGSO）， e.g.， C35 and C38. The results indicated that the fused prior was more reliable near the most significant bits （MSB） of the parameters and tended toward randomness near the least significant bits （LSB）. Setting the prior weight in the 0.4~0.8 range yielded a robust trade-off between low-SNR coding gain and unbiased decisions at high SNR. The results of an experimental evaluation showed that the proposed scheme reduced the bit error rate （BER） by roughly three to four orders of magnitude in the 1 to 2 dB SNR interval compared to a conventional EMS decoder. Moreover， it also achieved a decoding-threshold gain of about 1.0~1.5 dB at a target BER of 10^-5. Over a broader SNR range， the method remained stably convergent. Overall， it left-shifted the decoding threshold by approximately 1.5 dB compared with the baseline. These findings suggest that the fused-prior approach provides robust decoding gains under ultralow SNR conditions.