EMERGE
The official implementation of a CIKM 2024 paper that improves multimodal electronic health record prediction with retrieval-augmented summaries from a knowledge graph.
EMERGE extracts abnormal time-series signals and disease entities from clinical text, retrieves related disease knowledge from the PrimeKG knowledge graph, generates a patient-level retrieval-augmented summary, and fuses time-series, clinical-note and RAG-summary representations with cross-attention to improve multimodal EHR predictive modeling.
It expects EHR input in a three-table format covering dynamic time-stamped events, static patient attributes, and prediction labels, and ships a demo pipeline for the MIMIC-IV dataset that can be run end to end or stage by stage, covering download, format conversion, preprocessing and training. An optional extra installs BGE-M3 retrieval and Clinical-LongFormer embeddings for the more language-model-heavy parts of preprocessing.
The required PrimeKG files are downloaded automatically from the official Harvard Dataverse release, with an option to fetch the entire published PrimeKG dataset instead of only the files EMERGE itself needs; the loader accepts either the original .csv naming or the current Dataverse .tab naming for these files.