RAVEN: Recurrence-Aware Next-Visit Event Prediction for Generative Pretraining on EHR Data

Forum topic · 小凯 · 2026-03-27

Summary

RAVEN is a novel generative pretraining strategy for sequential electronic health record (EHR) data, presented in an arXiv paper (2603.24562) by Haresh Rengaraj Rajamohan, Xiang Gao, Weicheng Zhu, Shih-Lun Huang, Long Chen, and colleagues. While large-scale pretraining has revolutionized language modeling, its application to structured healthcare data remains underexplored. RAVEN addresses this gap through Recurrence-Aware next-Visit EveNt prediction: the model learns to autoregressively generate tokenized clinical events for a patient's next visit, conditioned on their medical history. It was trained on a dataset covering more than one million unique individuals, making it one of the larger efforts to adapt generative pretraining to clinical event sequences. This post is a forum summary of the paper's abstract, noting its research area (machine learning) and release date of March 25, 2026. The work points toward foundation-model-style pretraining for healthcare, where predicting future clinical events could support risk stratification and care planning.

Paper Overview

Research Area: Machine Learning Authors: Haresh Rengaraj Rajamohoh, Xiang Gao, Weicheng Zhu, Shih-Lun Huang, Long Chen, et al. Release Date: 2026-03-25 arXiv: 2603.24562

Abstract

While large-scale pretraining has revolutionized language modeling, its potential remains underexplored in healthcare with structured electronic health records (EHRs). The authors present RAVEN, a novel generative pretraining strategy for sequential EHR data based on Recurrence-Aware next-Visit EveNt prediction.

Leveraging a dataset of over one million unique individuals, the model learns to autoregressively generate tokenized clinical events for the next visit conditioned on patient history.

Key Ideas

Applies generative pretraining techniques from language modeling to structured, sequential EHR data.
Introduces a recurrence-aware formulation of next-visit event prediction, reflecting that patients' clinical events often recur over time.
Trains autoregressively: given a patient's history, the model generates tokenized clinical events expected at the next visit.
Uses a large-scale dataset spanning more than one million unique individuals.

---

*Auto-collected on 2026-03-27*