FHIR Terminology Encoder

This is a scikit-learn compatible encoder that uses a FHIR terminology server to encode ontological features.

It currently supports subsumption relationships and properties.

You supply a scope in the form of a FHIR ValueSet URI, and a FHIR terminology endpoint.

The result is a multi-hot encoded vector delivered as a sparse matrix, suitable for input into most models and estimators.

Installation

pip install fhir-tx-encoder

Usage

from fhir_tx import FhirTerminologyEncoder
import numpy as np

encoder = FhirTerminologyEncoder(
    # Ancestors of the SNOMED CT concept "Malignant neoplastic disease" (363346000)
    scope="http://snomed.info/sct?fhir_vs=ecl/(%3E%3E%20363346000)",
    # Include "Associated morphology" (116676008) as a property
    properties=["116676008"]
)

# Encode two SNOMED CT concepts:
# - "Neoplasm and/or hamartoma" (399981008)
# - "Malignant neoplastic disease" (363346000)
result = encoder.fit_transform(np.array([["399981008", "363346000"]]))

# Print out the result and its shape.
print(f"result.shape: {result.shape}")
print(f"result:\n{result.toarray()}")

# Print out the feature names.
print(f"encoder.feature_names_: {encoder.feature_names_}")

Which would output:

Expanding value set: http://snomed.info/sct?fhir_vs=ecl/(%3E%3E%20363346000)
Expanding (6 items, offset 0, total 6)
Expansion complete
Generating one-hot encoding... (6, 6)
Creating index... 6 items
Applying transitive closure...
Batch 1 of 1, 6 items... 15 pairs added
Subsumption encoding complete: (6, 6)
Encoding properties... (6, 9)
result.shape: (2, 9)
result:
[[1. 1. 0. 1. 0. 1. 0. 0. 1.]
 [1. 1. 1. 1. 1. 1. 0. 1. 0.]]
encoder.feature_names_: ['404684003', '64572001', '363346000', '399981008', '55342001', '138875005', '609096000.116676008=108369006', '609096000.116676008=1240414004', '609096000.116676008=400177003']

Important note

This software is currently in alpha. It is not yet ready for production use.

Copyright © 2023, Commonwealth Scientific and Industrial Research Organisation (CSIRO) ABN 41 687 119 230. Licensed under the Apache License, version 2.0.