NOAH: Learning the Full Patient Journey. A Longitudinal Multimodal Time-Aware Model for Representation and Forecasting
NOAH is a time-aware, task-agnostic, generative transformer model for representing and forecasting the full multimodal patient journey. It was built from over 559 million clinical events from 431,000 hospital visits of 299,000 patients across the MIMIC dataset family. NOAH processes medical images, time-series and numeric signals, and categorical events. The paper was published on arXiv on 2026-09-08.
The digitization of healthcare has generated vast, longitudinal, and multimodal patient records, but current AI models struggle to capture complex temporal dynamics and stochasticity. NOAH is introduced as a time-aware, task-agnostic, generative transformer model that represents and forecasts the full multimodal patient journey. It features a novel bidirectional time integration and a variational latent space to capture continuous patient state evolution and clinical trajectory stochasticity. The model was trained on over 559 million clinical events from 431,000 hospital visits of 299,000 patients across the MIMIC dataset family and natively processes medical images, time-series, numeric signals, and categorical events.
NOAH employs a bidirectional time integration mechanism and a variational latent space, enabling it to model irregular temporal dynamics and inherent stochasticity in multimodal patient data. Unlike prior discriminative models limited to few modalities and closed vocabularies, NOAH is generative and task-agnostic, capable of forecasting future patient states across diverse data types.
This research addresses a critical gap in healthcare AI by providing a unified model for longitudinal patient data. If validated, it could enable more accurate patient trajectory forecasting, personalized treatment planning, and early intervention, potentially reducing costs and improving outcomes in clinical settings.
NOAH could create value for healthcare providers, insurers, and pharmaceutical companies by improving risk stratification, resource allocation, and drug development through better patient journey modeling. It may also spawn new AI-driven clinical decision support products.
Next observable signals include peer-reviewed validation, clinical pilot studies, and potential open-source release of model weights or code. Adoption by healthcare institutions and integration with electronic health record systems would indicate progress toward real-world deployment.