NOAH: Learning the Full Patient Journey. A Longitudinal Multimodal Time-Aware Model for Representation and Forecasting

Researchers introduce NOAH, a task-agnostic, time-aware generative transformer trained on over 559 million MIMIC clinical events from 431,000 hospital visits by 299,000 patients to represent and forecast multimodal patient trajectories.

NOAH is a task-agnostic, time-aware generative transformer designed to represent and forecast the full multimodal patient journey. It was trained on over 559 million clinical events from 431,000 hospital visits covering 299,000 patients across the MIMIC dataset family. The architecture combines bidirectional time integration with a variational latent space to capture continuous patient state evolution and clinical stochasticity, and it natively processes medical images, time-series signals, categorical events, and structured or unstructured clinical records. NOAH supports autoregressive forecasting with optional time control, zero-shot classification, and counterfactual intervention simulation. The two source reports agree on all figures; the Hugging Face report describes strong probing performance across clinical outcomes, 15 ICD chapters, and 29 comorbidities, while the arXiv report frames the same evaluations as spanning 15 ICD chapters, 29 comorbidities, and time-to-event prediction. No discrepancies between sources were identified.

• NOAH is a task-agnostic, time-aware generative transformer for longitudinal multimodal patient records (stated in both reports).
• Trained on over 559 million clinical events from 431,000 hospital visits by 299,000 patients across the MIMIC dataset family.
• Architecture combines bidirectional time integration with a variational latent space to model stochastic evolution of patient states.
• Natively handles medical images, time-series signals, categorical events, and structured or unstructured clinical text.
• Supports autoregressive forecasting with optional time control, zero-shot classification, and counterfactual intervention simulation.
• Evaluated/probed across 15 ICD chapters and 29 comorbidities; the arXiv report additionally cites time-to-event prediction evaluations.
• Source reports: Hugging Face daily papers (2026-09-07T20:00:00Z) and arXiv cs.AI/cs.LG/cs.CL (2026-09-08T17:56:13Z); all cited figures are consistent across both.