Researchers have introduced NOAH, a task-agnostic, generative transformer model designed to represent and forecast longitudinal, multimodal patient trajectories. To address the limitations of existing discriminative models—such as irregular temporal dynamics and closed vocabularies—NOAH utilizes a novel bidirectional time integration mechanism and a variational latent space. This architecture allows the model to capture both the continuous evolution of patient states and the inherent stochasticity of clinical data. The model was trained on a massive dataset comprising over 559 million clinical events from 431,000 hospital visits across 299,000 patients within the MIMIC dataset family. NOAH is uniquely capable of natively processing diverse modalities, including medical images, time-series/numeric signals, categorical events, and both structured and unstructured clinical records. As a holistic generative model, it supports autoregressive forecasting with time control, zero-shot classification, and counterfactual intervention simulations. Experimental results indicate that NOAH generates highly informative representations that excel in probing clinical outcomes, 15 ICD chapters, and 29 different comorbidities, while also demonstrating strong performance in time-to-event prediction tasks. This framework provides a scalable foundation for personalized clinical care by seamlessly integrating complex temporal dynamics with diverse medical data types.