This report is an investigation of an alternate method for inference in a large graphical BN2O model used by the AI system IntMed. The system is used for clinical decision support in real-time medical consultations and is made by Ambolt ApS who have supplied the data necessary to conduct the research. The current inference calculation algorithm of IntMed is called 'Quickscore' and calculates exact probability estimations, but struggles to uphold the real-time constraint. We propose the usage of a special type of neural networks known as Recognition Networks for inference approximation. Specifically, we propose a Recurrent Recognition Network capable of analysing the temporal unveiling of symptoms through patient questioning that happens during consultations. We show how this recurrent network can be trained using forward sampling from a BN2O. To demonstrate the network's potential we compare it to Quickscore in various consultation scenarios, in terms of posterior estimation and importance order of diseases. These results show that a recurrent neural network is able to mimic the results of exact inference better than the sequential counterpart, but needs attention in terms of scaling and calibration.