OHS Analytics Concepts

The key concepts that underpin the OHS Analytics components are:

ETL Pipelines: ETL Pipelines and Controller can be configured to continuously transform FHIR data into an analytics friendly FHIR-in-Parquet format.
Deployment approaches: The pipelines are designed to accommodate various deployment approaches in terms of scalability; from a single machine to a distributed cluster environments.
Query simplification approaches: Once the data is transformed into an analytics-friendly format, it should be queried. Multiple approaches are provided and/or showcased to flatten FHIR schema to make developing analytics solutions easier.

FHIR Data Pipes

FHIR Data Pipes is built on Apache Beam SDK for ease of scalability and has multiple deployment options from local, to on-prem clusters to public clouds.

FHIR Data Pipes is made up of the ETL Pipelines and Controller modules * that are designed to work together* to provide continuous transformation of FHIR data to Apache Parquet files (for data analysis) or another FHIR server ( for data integration).

FHIR Data Pipes Transform Step Image

ETL Pipelines

Java binaries to Extract, Transform and Load FHIR data from a FHIR source to Parquet format.

Extraction

FHIR Data Pipes is designed to fetch FHIR source data in various forms and APIs:

FHIR Search API: This should work with most FHIR servers, including those using FHIR data model like HAPI FHIR or FHIR facades like OpenMRS.
Direct database access: This is mostly a more efficient approach that works with HAPI FHIR only.
Bulk Export API: This will gradually become our main fetch API as more FHIR servers implement the Bulk Export API.
Files in json and ndjson formats.

Transformation

FHIR Resources are transformed into a "Parquet on FHIR" format:

Uses a forked version of Bunsen library (_ currently supports STU3 and R4 versions of FHIR).
Configurable support for FHIR profiles and extensions
(Optional) In-pipeline 'flattening' of FHIR data using ViewDefinition resources - read more

Loading

FHIR Data Pipes supports different SQL Data Warehouse options depending on the needs of the project. These include:

Loading Parquet files into an OLAP query engine such as SparkSQL (which can be deployed on a single machine or a distributed cluster) or DuckDB (single machine in-process) or many other tools that understand Parquet.
Traditional relational databases such as PostgreSQL (when using FHIR ViewDefinition resources to generate materialized views)

Pipeline Controller Module

A user-interface wrapper for the FHIR Data Pipes Pipelines, integrating " full", "incremental", and "merger" pipelines together.

The Pipelines Controller is built on top of pipelines and shares many of the same settings
Using the controller module you can schedule periodic incremental updates or use a web control panel to start the pipeline manually

Deployment approaches

There are a number of different deployment approaches - see table below.

Choosing the right approach, comes down to a number of factors including the specific use-case, infrastructure constraints of the project, future scalability requirements, and expertise of the team.

Scenario	Approach	Considerations
Simple relational database to power dashboards or reporting	Custom schema defined as ViewDefinition Resources - see tutorial	By design, this will provide a subset of FHIR data in flat tables
Horizontally scalable query architecture with lossless FHIR data	Parquet based DWH and distributed query engine (e.g. SparkSQL) - see tutorial	This requires a distributed file system for Parquet files and a distributed query engine.
Lossless FHIR DWH with a single process query engine	Parquet based DWH with non-distributed query engine set-up (e.g. single-node Spark or DuckDB)	Scalability and performance (e.g. SparkSQL >> DuckDB)
Exploratory data science or ML use cases	Use the generated Parquet files which as "near lossless" for enhanced data science workflows	Can either use the Parquet or custom schema to power dashboards or reports
Push FHIR data to a central FHIR-store (e.g., for a Shared Health Record system)	Use the Pipelines Controller to push from a FHIR source to a FHIR sink	Management of the intermediate Parquet files created as part of the pipelines

Query simplification approaches with pre-defined views

The heavily nested nature of FHIR resources requires complex SQL queries that can make it difficult to work with for analytics use cases. A common approach to address this is to flatten the data into a set of views (virtual or materialized) which can then be queried using simpler SQL statements.

FHIR Data Pipes provides two approaches for flattening the FHIR resources into virtual or materialized views:

SQL queries to generate virtual views (outside the pipeline) 2. FHIR ViewDefinition resources to generate materialized views (within the pipeline)

For both of these approaches, a set of **"predefined views” ** for common resources are provided. These can be modified or extended.

The currently supported list (as of June, 2024) are:

Condition
DiagnosticReport
Encounter
Immunization
Location
Medicationrequest
Observation
Organization
Patient
Practitioner
PractitionerRole
Procedure

SQL virtual views

These are samples of more complex SQL-on-FHIR queries for defining flat views for common FHIR resources. These virtual views are applied outside of the pipelines in the downstream SQL query engine.

The queries, which have .sql suffix, can be found in /docker/config/views directory (e.g Patient_flat.sql).

An example of a flat view for the Observation resource is below:

CREATE OR REPLACE VIEW flat_observation AS
SELECT O.id AS obs_id, O.subject.PatientId AS patient_id,
        OCC.`system` AS code_sys, OCC.code,
        O.value.quantity.value AS val_quantity,
        OVCC.code AS val_code, OVCC.`system` AS val_sys,
        O.effective.dateTime AS obs_date
      FROM Observation AS O LATERAL VIEW OUTER explode(code.coding) AS OCC
        LATERAL VIEW OUTER explode(O.value.codeableConcept.coding) AS OVCC

ViewDefinition resource

The SQL-on-FHIR-v2 specification defines a ViewDefinition resource for defining views. Each column in the view is defined using a FHIRPath expression. There is also an unnesting construct and support for constant and where clauses too.

A system (pipeline or library) that implements the “View Layer” of the specification provides a View Runner that is able to process these FHIR ViewDefinition Resources over the “Data Layer” (lossless representation of the FHIR data). The output of this are a set of portable, tabular views that can be consumed by the “Analytics Layer” which is any number of tools that can be used to work with the resulting tabular data.

FHIR Data Pipes is a reference implementation of the SQL-on-FHIR-v2 specification:

The "View Runner" is by default part of the ETL Pipelines and uses the transformed Parquet files as the “Data Layer”. This can be extracted to be a stand-alone component if required
When enabled as part of the Pipeline configuration, it will apply the ViewDefinition resources from the views folder and materialize the resulting tables to the configured database (e.g., an instance of PostgresSQL, MySQL, etc.).
A set of pre-defined ViewDefinitions for common FHIR resources is provided as part of the default package. These can be adapted, replaced and extended.
The FHIR Data Pipes provides a simple ViewDefinition Editor which can be used to explore FHIR ViewDefinitions and apply these to individual FHIR resources.

Once the FHIR data has been transformed via the ETL Pipelines, the resulting schema is available for querying using a JDBC interface.

ViewDefinition editor

The ViewDefinition editor provides a way to quickly evaluate ViewDefinition resources against sample FHIR data. You access it as part of the Web Control Panel, selecting the "Views" navigation item in the top right corner.

Using the ViewDefinition editor you can:

Provide an input ViewDefinition (left)
Apply it to a sample input FHIR resource (right pane)
View the results in the generated table (top)

FHIR Data Pipes Image