Algorithmic mapping

The main recommended algorithmic techniques applicable to wider terminology mapping are:

• semantic and lexical mapping
• using terminology component building blocks
• configurable mapping to additional data elements that supplement source data
• using configurable terminology translation algorithms
• using referential terminology mapping artifacts

They are exemplified and specified via the pathology and laboratory medicine terminology mapping use-case in subsequent subsections. These are the fundamental techniques the Pathology and Laboratory Medicine (PaLM) Mapping Project team investigated to optimise automated terminology mapping outputs during testing.

This section describes how these techniques can be used to help you map single reportables to SNOMED CT PaLM (that is, mapping terminology descriptions pertaining to a single code between two terminology systems). However, the same underlying techniques can facilitate both mapping to supplemental data elements in the Pathology FHIR Specification and identifying multiple SNOMED CT PaLM reportables for complex microbiology reports.

Most mapping tools use string based matching algorithms to establish map targets (see the Advanced Analysis Components to Support SNOMED PaLM Mapping Project paper for details). 

In this example, the local reportable can easily be mapped to the SNOMED CT PaLM target due to the close similarity between source and target human-readable descriptions.

SNOMED CT PaLM reportables carry more than one human-readable description: the ‘fully specified name’, the ‘preferred term’ (which often includes acronyms), and in some instances, additional synonyms. All these should be used to facilitate mapping.

Acronyms are prevalent in local reportable descriptions, so the ability of mapping tools and large language models (LLMs) to process acronyms is particularly valuable.

To assist review, a mapping tool can establish the level of equivalence between terms being matched (such as equivalent, more specific, less specific), thereby offering a level of confidence in the validity of the proposed map.

These techniques use the strategy defined in Section 3 - mapping strategies.

Configurable mapping to additional data elements that supplement source data

This example shows the source data relating to a lab’s local ‘serum creatinine’ reportable.

Note -  umol/L defines the property as ‘a ‘substance concentration’

Configuring an algorithm, then semantically/lexically mapping these source data elements to SNOMED CT PaLM reportables’ descriptions, or to the descriptions of the SNOMED CT concepts used in their modelling, will establish the matching SNOMED CT PaLM reportable.

Here is an example:

Using configurable terminology translation algorithms

During testing, the PaLM Mapping Project team created a map table of common units of measure defined as lab test properties used in the modelling of SNOMED CT PaLM, as per the example below: 

Whilst not a clinically assured artefact, this table was used in testing as a referential artefact to facilitate mapping. Testing found that applying an algorithmic translation of units of measure to lab test property enabled the mapping of large volumes of reportables. When combined with a specimen type extracted via labs’ source data or from the mapped Read PBCL description, this became a powerful method of identifying SNOMED CT PaLM map targets. 

This section explores how loading referential terminology concept maps into a tool can greatly enhance the automated mapping output. 

Using other labs’ curated maps to SNOMED CT PaLM

Where one lab's local reportables have been mapped to SNOMED CT PaLM, that information can help establish a second lab's mapping, especially when existing maps to Read PBCL/SNOMED CT PBCL are factored in.

In this example, Hospital A and Hospital B have mapped their local reportables to the same Read PBCL/SNOMED CT PBCL code. Hospital A are unsure which SNOMED CT PaLM reportable to map to, whereas Hospital B have established that their SNOMED CT PaLM map is Qualitative result of Salmonella species culture (observable entity). Consequently, Hospital B’s map could be used to evaluate whether the same SNOMED CT PaLM reportable is appropriate for Hospital A.

Using local reportable to Read PBCL maps 

Labs reporting results to general practice all use their own local reportable to Read PBCL maps to support the PMIP EDIFACT flow. The NHS England assured mapping table from Read PBCL to SNOMED CT PBCL is then used by GP systems to ingest these reports and record lab test result codes as SNOMED CT in the patient record.

Whilst there is a degree of variance in the quality of labs’ existing Read PBCL maps, they remain a vital resource to facilitate mapping to SNOMED CT PaLM. Using the information contained in these maps can help in three ways.

1. Using Read PBCL/SNOMED CT PBCL to SNOMED CT PaLM maps 

During testing, the PaLM Mapping Project team created a map table between Read PBCL/SNOMED CT PBCL and SNOMED CT PaLM.

This map table was generated by repurposing a tool created to support the modelling of SNOMED CT PBCL and SNOMED CT PaLM. It uses a bespoke set of semantic parsing instructions and a MySQL database.  

The diagram below shows the process involved. Each SNOMED CT PBCL and SNOMED CT PaLM reportable’s human-readable descriptions are split into component parts (as per the example given in section 6 - Parsing terminology into component building blocks). Algorithms were then applied against SNOMED CT attribute values used in the modelling of SNOMED CT reportables to identify appropriate map targets. These were then scored in terms of appropriate matching.

Whilst not a clinically assured artifact, these concept maps were used in testing as a referential terminology mapping artefact to facilitate mapping. It was achieved by using labs’ local reportable to Read PBCL concept maps as a primary source input. The Read PBCL codes were subsequently mapped to SNOMED CT PBCL reportables, which in turn were mapped to SNOMED CT PaLM targets, as seen in the diagram below.

The process is explored further in section 8 - Managing the mapping process.

2. Establishing source data via Read PBCL/SNOMED CT PBCL maps

As previously explored (section 6 - Techniques for source data cleansing), Read PBCL/SNOMED CT PBCL maps provide a means to establish source data.

In this example, the local reportable ‘PLT’/‘Platelets’ has been mapped to Read PBCL/SNOMED CT PBCL. The additional word ‘count’ in the Read PBCL and SNOMED CT PBCL descriptions establishes the lab test’s property. This additional information can be used to semantically match to the SNOMED CT PaLM reportable.

Note - testing found this particularly powerful in helping to help determine specimen types. 

3. Using SNOMED CT taxonomic relationships

In this example, the local reportable ‘NA’/‘Sodium’ has been mapped to Read PBCL and the equivalent SNOMED CT PBCL reportable 'Serum sodium level (observable entity)'.

Via the hierarchical SNOMED CT ‘Is a’ relationship, the SNOMED CT PaLM reportable 'Substance concentration of sodium in serum (observable entity)' is a ‘child’ of 'Serum sodium level (observable entity)', as seen in the image below:

Consequently, it can be established that the SNOMED CT PaLM concept is a potentially suitable map target via algorithms that use SNOMED CT’s logical concept model. This is particularly useful as a secondary check to compliment other mapping techniques.

This diagram illustrates the various data elements contained in a DAPB4101 conformant Neisseria gonorrhoeae DNA detection test report, as defined by the Pathology FHIR specification. The SNOMED CT PaLM reportable Qualitative result of Neisseria gonorrhoeae deoxyribonucleic acid nucleic acid amplification (observable entity) represents the lab test result code.

Reportables such as these do not inherently carry information about the specimen and lab test technique. To ensure this coded information is carried in the report, lab source data representing the specimen and technique can be mapped to SNOMED CT Specimen and Procedure concepts respectively. Likewise, lab source data representing the result type can be mapped to an equivalent SNOMED CT Qualifier value concept. 

Although the mapping techniques remain the same, the process involves creating target value sets of applicable SNOMED CT concepts. This necessitates access to wider SNOMED CT content than the PaLM (Pathology and Laboratory Medicine) observable entity simple reference set. Accessing this content is similar to the process described in Section 5 - How to access SNOMED CT PaLM, and likewise, mapping tools can facilitate this. Dependent on the tool’s functionality, constraining map targets may require the use of simple ECL expressions, such as < 123038009 |Specimen (specimen)|.

As described in section 5 - Single reportables vs 'groupers', SNOMED CT PaLM reportables are single reportables, and multiple SNOMED CT PaLM reportables are contained in complex microbiology reports.

The diagram below illustrates the various data elements contained in a DAPB4101 conformant urine culture and sensitivities complex report, and includes several SNOMED CT PaLM reportables. The challenge for labs is to identify which SNOMED CT PaLM reportables are appropriate for inclusion. 

This can be done by:

1. Mapping ancillary lab source data used in the report to SNOMED CT concepts used in the modelling of SNOMED CT PaLM reportables (these concepts being the reportables’ ‘attribute values’).
2. Applying algorithms that use SNOMED CT’s concept model to identify which SNOMED CT PaLM reportables use this particular modelling, thereby identifying them as appropriate for inclusion. 

In this example, lab source data representing the organism and result type can be mapped to SNOMED CT 'Organism' and 'Qualifier value’ concepts. This helps to identify the appropriate SNOMED CT PaLM reportable for the culture and sensitivities section of the microbiology report.

Lab source data representing the Staphylococcus aureus organism can be mapped to the SNOMED CT concept Staphylococcus aureus (organism) which is used as the ‘component’ attribute value in the modelling of several SNOMED CT PaLM reportables.

Lab source data representing the result value ‘scanty growth’ can be mapped to the SNOMED CT concept Scanty growth (qualifier value). This result value identifies the reportable as a Qualitative result type.

An algorithm can combine these two pieces of information to identify the appropriate SNOMED CT PaLM reportable:

Likewise, lab source data representing antimicrobials can be mapped to SNOMED CT 'Medicinal product' concepts to help identify the appropriate SNOMED CT PaLM reportable for the antimicrobial sensitivity section of the microbiology report.

Lab source data representing the antimicrobial flucloxacillin can be mapped to the SNOMED CT concept Product containing flucloxacillin (medicinal product). This is used as the ‘towards’ attribute value in the modelling of the SNOMED CT PaLM reportable Susceptibility of organism to flucloxacillin (observable entity):