Core Concepts
RAM-EHR enhances clinical predictions on EHRs by leveraging dense retrieval and multiple knowledge sources, leading to improved model performance.
Abstract
RAM-EHR introduces a novel approach to enhance clinical predictions on Electronic Health Records (EHRs) by utilizing dense retrieval and multiple knowledge sources. The system collects diverse knowledge sources, converts them into text format, and uses dense retrieval to obtain information related to medical concepts. By augmenting the local EHR predictive model with summarized knowledge, RAM-EHR achieves significant gains in AUROC and AUPR over previous baselines. The co-training approach captures complementary information from patient visits and external knowledge, improving generalization and model performance. Experiments on two EHR datasets demonstrate the effectiveness of RAM-EHR in enhancing clinical prediction tasks.
Stats
RAM-EHR shows a 3.4% gain in AUROC and a 7.2% gain in AUPR over previous baselines.
Multiple knowledge sources are utilized for enhanced predictions.
Co-training with consistency regularization improves model generalization.
Quotes
"RAM-EHR offers flexibility and can seamlessly integrate diverse sources of knowledge."
"Experiments confirm the advantage of leveraging multi-source external knowledge for clinical tasks."
"Our contribution introduces an innovative framework designed to harness external knowledge for enhanced clinical predictions."