Learning-Based Semantic Alignment in Cross-Institution Health Data Exchange

Abstract

Healthcare ecosystems today are built upon the exchange of health data between multiple types of organizations, including provider organizations, payers, laboratories, public health agencies, and government programs. While syntactical standards, such as HL7 v2, FHIR, X12, and CDA have been widely adopted to enable the exchange of data electronically, the lack of consistent semantics across institutions continues to create barriers to the correct interpretation of data exchanged electronically, automated workflows and analytics. The variability in schema design, terminology use, contextual meaning, and workflow specific customization creates ambiguity in data exchanged electronically which cannot be eliminated by using rule based mapping techniques or static ontologies. In this paper we present a novel, learning-based semantic alignment framework for cross-institutional health data exchange which utilizes representation learning, contextual embeddings and probabilistic alignment scoring to align schematically disparate healthcare datasets. Our proposed approach consists of three main components; schema-level feature learning to extract common features from the source and target schemas, a context aware similarity model that measures the similarity between the two schemas, and an explainability driven validation process to ensure that the alignment decisions made are trusted and suitable for regulated environments. We also present a multilayer architecture to facilitate both semantic discovery, alignment confidence estimation, human-in-the-loop review and regulatory auditability. We evaluate our proposed approach on synthetic but realistic healthcare schemas developed from electronic health record (EHR) data, claims data and enrollment system data, and compare our approach with two baseline approaches; one that uses rules to align the two schematics and another that uses an ontology to align the two schematics. Finally, our proposed approach includes governance controls that meet all applicable requirements of HIPAA, CMS, and NIST AI Risk Management Framework (RMF).