TL;DR
This paper introduces Archetype Relational Mapping (ARM), a novel approach to storing complex electronic health record (EHR) data in relational databases. ARM addresses the challenge of adapting EHR systems to evolving healthcare data by mapping openEHR archetypes to relational tables, achieving performance comparable to traditional EHR systems and facilitating the adoption of the openEHR architecture.
What Problem Does This Paper Address?
The widespread adoption of openEHR, a standard for future-proof EHR systems, is hindered by the lack of practical persistence solutions. Traditional EHR systems often use rigid, single-level data models that struggle to accommodate the dynamic and complex nature of healthcare information. openEHR's two-level modeling approach, using a stable Reference Model (RM) and flexible archetypes, offers a solution, but mapping these archetypes to efficient and adaptable relational databases remains a significant challenge. Existing solutions like Node+Path have limitations in performance and scalability. This paper aims to bridge this gap by proposing ARM, a method to generate relational database schemas directly from openEHR archetypes and templates, ensuring both flexibility and performance.
Methods
The Archetype Relational Mapping (ARM) approach involves three main steps:
- Archetype Modelling: This involves selecting existing archetypes from the Clinical Knowledge Manager (CKM) or defining new ones to represent healthcare data concepts. The goal is to create archetype-friendly concepts that capture detailed data requirements.
- Template Definition: Templates are used to adapt archetypes to specific local EHR contexts. They define constraints, such as local optionality and tightened data constraints, and importantly, ARM constraints are designed to align the archetype model with the relational database model for better performance. Key constraints include assigning identification and query data items, and defining mappings between generalized and specialized archetypes.
- Database Mapping: This is the core of ARM, where archetypes and templates are translated into a relational database schema. Rules are defined to map each archetype to a table, basic data items to columns, and to handle complex structures like archetype slots and collection data items. The process aims to create a database that can automatically persist data represented by archetype instances.
Datasets
The study evaluated the ARM approach using data from a tertiary Class A hospital in China. This data was extracted from an existing Integrated Viewer (IV) information system, which consolidates data from various legacy systems (HMIS, LIS, RIS, PACS, PMS, OMS, CPOE). The dataset included approximately 103,320 imaging tests, 857,315 images, 654,213 laboratory tests, and 484,668 laboratory test results for 29,743 patients. The data was de-identified before use. The study also compared ARM's performance against the original IV database and the openEHR official Node+Path database.
Key Insights and Ideas
- Bridging the Semantic Gap: ARM effectively translates the semantic richness of openEHR archetypes into a structured relational database, making complex healthcare data queryable and manageable within traditional database systems.
- Flexibility Meets Performance: The approach demonstrates that it's possible to achieve performance comparable to conventional EHR systems while retaining the flexibility offered by archetype-based modeling, a key challenge in EHR development.
- De-normalization for Speed: To optimize query performance, ARM employs strategic de-normalization techniques, such as embedding archetypes and propagating indices, which reduce the complexity and number of joins required for data retrieval.
- Template-Driven Adaptation: The use of templates is crucial for tailoring the generic archetypes to specific local EHR requirements, ensuring that the persistence layer is not only flexible but also practical for real-world clinical use.
Key Results
The performance comparison revealed that:
- Data Retrieval: ARM performed comparably to the conventional IV database in data-retrieving tests. ARM was better in tests 1, 2, and 4, while the IV database was slightly faster in tests 3 and 5. The time differences were generally within 130%.
- Patient Searching: ARM significantly outperformed the conventional IV database in patient-searching tests, being approximately 6–50 times more efficient. This is attributed to ARM's ability to map archetypes to standalone tables, avoiding the EAV-style storage of the IV database.
- Node+Path Inefficiency: The Node+Path database consistently performed much worse than both ARM and the IV database across all tests, requiring significantly more time, especially for sparse data and complex queries.
- Storage: The ARM database required 2.90 GB of storage, while the conventional IV database required 1.60 GB, and the Node+Path database required a substantial 43.87 GB.
Limitations and Future Work
The authors acknowledge several limitations:
- Highly-Tuned Scenarios: ARM might not meet the requirements of situations demanding extremely high database tuning, as index optimization is complex and context-dependent.
- Archetype Availability and Quality: The effectiveness of ARM relies heavily on the availability and quality of published archetypes. If local data requirements differ significantly from existing archetypes, extensive effort is needed to develop new, high-quality archetypes.
- De-normalization Complexity: While beneficial for performance, de-normalization can introduce complexity in data querying, requiring careful consideration of how to query embedded archetypes.
Future work suggested includes:
- Continuing to refine mapping rules based on archetype semantics.
- Improving ARM constraints and the performance of generated databases.
- Designing and implementing data access services.
- Conducting more thorough tests of the ARM approach in real clinical environments.
Summary updated