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

Imagine you're a healthcare provider trying to understand a patient's journey through the healthcare system. You have access to a wealth of data — medical images, lab results, and clinical notes — but piecing it all together to predict future health outcomes is a daunting task. Current AI models often fall short because they focus on specific types of data or treat time in a simplistic way, which can lead to missed insights and inaccurate forecasts. This is what's called a lack of holistic understanding of patient dynamics. They struggle with the irregularities and complexities of real-world data, making it hard to provide personalized care based on a patient's unique history and future risks.

To address these challenges, NOAH offers a fresh approach by using a generative transformer model that can handle diverse data types and capture the intricate temporal relationships in patient records. It integrates information bidirectionally over time and employs a variational latent space to model the continuous evolution of patient states. This means it can not only forecast future health outcomes but also simulate different clinical scenarios and classify conditions without needing extensive retraining. Compared to previous models, NOAH provides a more comprehensive and adaptable framework for understanding patient trajectories, which could significantly enhance predictive systems in personalized medicine and improve patient care outcomes.

NOAH introduces a novel generative approach to modeling complex patient trajectories across multiple modalities.

The model is built on a large dataset and demonstrates strong performance across various clinical tasks, though specific baselines are not detailed.