The integration of Electronic Health Records has transformed patient data management in healthcare, offering comprehensive insights into patient journeys through diverse data types. Despite the value of these records for developing diagnostic models, their use is heavily regulated by strict regulations to ensure patient privacy and confidentiality. Generative deep learning models can generate realistic synthetic data to circumvent these restrictions. Patient trajectories, marked by variable lengths and multi-modality, include irregularly sampled medical time series, which is often challenging for models to handle. In this thesis, we aim to create a GAN-based architecture that is able to handle as well as mimic the variable length and multi-modality of patient trajectories. To evaluate the synthetic dataset, we used three types of evaluation metrics based on fidelity, utility and privacy. Our results indicate that while the proposed model outperforms the baselines in some evaluation tests, it performs subpar in others,concluding that the fundamental ideas work and further improvements are possible.