NOAH: Learning the Full Patient Journey. A Longitudinal Multimodal Time-Aware Model for Representation and Forecasting
NOAH models a patient record as a time-aware, multimodal sequence it can both represent and forecast.
The paper presents a generative transformer trained on more than 559 million clinical events from the MIMIC dataset family, covering 299,000 patients and 431,000 hospital visits. It is designed to handle images, time series, numeric signals, categorical events, and structured or unstructured clinical records in one patient-journey model. The authors say its bidirectional time integration and variational latent space are meant to capture irregular clinical timing and uncertainty. They report use cases including autoregressive forecasting, optional time control, zero-shot classification, counterfactual intervention simulation, and prediction probes for outcomes, ICD chapters, comorbidities, and time-to-event tasks. HF Daily Papers' note
The paper presents a generative transformer trained on more than 559 million clinical events from the MIMIC dataset family, covering 299,000 patients and 431,000 hospital visits. It is designed to handle images, time series, numeric signals, categorical events, and structured or unstructured clinical records in one patient-journey model. The authors say its bidirectional time integration and variational latent space are meant to capture irregular clinical timing and uncertainty. They report use cases including autoregressive forecasting, optional time control, zero-shot classification, counterfactual intervention simulation, and prediction probes for outcomes, ICD chapters, comorbidities, and time-to-event tasks. HF Daily Papers' note
score 4