A Gaussian matrix graphical encoder in sports medicine diagnosis combining structured and unstructured data
Abstract: We study the integration of Electronic Medical Records (EMRs) from clinical study into a joint predictive model. Compared to the totally black-box models, a competitive model with explainable structure is much more desirable. To tackle this challenge, this paper introduces
a novel Gaussian Matrix Graphical Encoder(GMGE) based on matrix normal graphical model to encode unstructured medical text and simultaneously learn the underlying conditional dependency graph of concepts. We further present DiMES, a Diagnostic Model with Explainable Structure, which integrates the concept graph generated by GMGE with structured data such as patient's physical examination measures. Utilizing Graph Convolutional Networks (GCNs), DiMES encodes patient features based on the concept graph for downstream tasks, providing clinicians with accurate predictive information to assist in diagnostic decisions and treatment plan design. The effectiveness of the proposed DiMES is validated through its application on four downstream diagnostic predictive tasks(ACL, PCL, MMI and PS).
Paper Type: Long
Research Area: NLP Applications
Research Area Keywords: Gaussian matrix graphical model, encoder, sport medicine, unstructured data, structured data
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings, Data analysis
Languages Studied: English, Chinese
Submission Number: 5942
Loading