As discussed in the Section about Substrates for Analytics, not all electronic health records represent clinical data using SNOMED CT. However, even when health records use free text or other code systems, it is still possible to use SNOMED CT for analytics over this data if the data is prepared appropriately. For example, Natural Language Processing can be used to encode free text data in SNOMED CT, subsequently enabling more sophisticated analytics to be performed. Similarly, clinical data using other code systems can be mapped into SNOMED CT to provide similar benefits.

In this section, we discuss these alternative ways of preparing clinical data for analytics using SNOMED CT.

While there is a strong trend towards the direct coding of clinical data, the capture and retention of free text remains essential to record broader narratives about clinical history, physical examinations, clinical procedures and investigation results. Wider deployment of medical transcription technologies featuring speech recognition also add to the volume of free text in electronic format. Medical literature, clinical guidelines and published clinical research also remains largely in free text.

Natural Language Processing (NLP) is a linguistic technique that enables a computer program to analyze and extract meaning from human language. Clinical NLP, using SNOMED CT's concepts, descriptions and relationships, may be applied to repositories of clinical information to search, index, selectively retrieve and analyze free text. These techniques can be used to extract SNOMED CT encoded data from free-text patient records, and also support the retrieval of clinical knowledge documents.

It should be noted that while clinical NLP techniques have increased in sophistication over recent years, it is not possible to guarantee full accuracy or completeness using a computer-based algorithm. Spelling errors, grammatical errors, abbreviations, unexpected synonyms, unusual vernacular (i.e. local) phrases, and hidden contextual information continue to provide challenges that human intelligence is uniquely equipped to handle.

Example

The example shown below in Figure 5.1-1 shows a free text section of a discharge summary that has been processed with clinical NLP to extract a set of potential SNOMED CT clinical findings and procedures. In order to ensure the correctness of this automatic encoding, the application should present this list of extracted codes to the user for confirmation, giving them the opportunity to refine, delete or append codes.

To improve the accuracy of clinical NLP and the value for analytics processes, it is important that the context of each statement expressed in natural language is clearly identified – for example, past history, suspected and negation/absence. shows the same discharge summary narrative as in , but this time processed with clinical NLP that also extracts the explicit context of each clinical finding and procedure.

When SNOMED CT codes with explicit context are extracted from free text narrative, the resulting clinical meanings may be captured using SNOMED CT postcoordinated expressions. For example, the following clinical statement:

Endoscopy revealed an acute gastric ulcer but no evidence of gastric bleeding or perforation of the stomach.

can be encoded using the following SNOMED CT expressions with explicit context (see case study):

A clinical NLP engine can use SNOMED CT to encode free text narrative in patient records in a number of ways. Firstly, it can use SNOMED CT descriptions together with techniques such as:

• Stemming: The process of reducing a word to its stem, base or root form – for example "cardiology", "cardiac" and "cardiologist" may be reduced to the stem "cardi".

• Reordering: The process of reordering the words in a phrase – for example, reordering "hip fracture" to "fracture hip".

• Word substitution: The process of substituting a word or word phrase with an equivalent word or word phrase. The SNOMED CT Lexical Resources zip file, available from the SNOMED CT Document Library, includes an English Word Equivalents table that groups together equivalent words and phrases – for example, "Renal stone", "Kidney stone", "kidney calculus", "renal calculus" and "nephrolith" are grouped into the same word block group. This table can be modified or extended with additional word equivalent groups if required.

The SNOMED CT concept model can also be used to identify potential connections between related concepts – for example, the words "left", "hip" and "fracture" used in close proximity may indicate a |fracture| with finding site |hip| and a laterality of |left|. Similarly, the SNOMED CT concept model may help to identify context that is expressed within the text – for example, past history, certainty and absence.

Another commonly adopted NLP strategy is to use the location of the free text within the structure of a document to restrict the possible SNOMED CT code matches. For example, free text entered into a 'Diagnosis' field may restrict its SNOMED CT encoding to the |disorder| hierarchy, together with other concepts that may be linked to |clinical findings| via the SNOMED CT concept model.

When NLP techniques are applied to non-English (or dialect-specific) text, translations of relevant SNOMED CT descriptions may be required. The NLP methods themselves may also need to be adapted to reflect the structure and style of the language in which the text is written.

Another major application for Natural Language Processing technologies is indexing collections of free text transcripts or documents such that topic specific searches may be run on them, or relevant clinical knowledge sources may be identified and linked to a given patient's clinical data. The challenge is to return ranked matches which permit selection of texts with high sensitivity and high specificity (i.e. that relevant documents are rarely overlooked and that irrelevant documents are rarely returned).

SNOMED CT can be used to support these applications by enabling more powerful searching of free text data stores than using a purely lexical keyword-based approach. For example, the clinician may request "all documents which refer to cardiac rhythm disorders". Rather than relying purely on text matching, the search term may be matched with the concept 698247007 |cardiac arrhythmia (disorder)|, based on its synonym |disorder of heart rhythm|. The descendants of this concept (e.g. 276796006 |atrial tachycardia|, 49260003 |idioventricular rhythm|, 233917008 |atrioventricular block|) may then be used to search for any code which is a kind of cardiac arrhythmia. Non-|is a| attribute relationships may also be used in the retrieval process to find associations between the search term and the indexed concepts, and to calculate the relevance of each free text artefact to determine the order in which they should be presented to the user.

Clinical NLP has been implemented for encoding free text narrative in health records by a number of vendors.

NLP techniques for indexing and searching have also been implemented by Cerner and Dr Bevan Koopman. Allscript's Sunrise InfoButton™ feature uses encoded patient problem lists and medication data elements, together with SNOMED CT-based indexes provided by third-party medical content providers, to present on-topic information to the clinician without manual searching.

Mapping data from clinical records encoded using non-SNOMED CT code systems to SNOMED CT for analysis may be considered when there is a requirement to produce:

• Management information for care service audit or delivery planning

• Statistical information for epidemiology

• Links from clinical records to clinical knowledge resources

• Links between clinical records and decision support tools

• An integrated data warehouse for querying from multiple heterogeneous sources

• Other types of research, reports or surveillance that requires SNOMED CT

Two important characteristics of a map, which affect its ability to be used for a particular purpose, are the direction of the map, and the correlation between the source and target codes. Where the analytics use case requires SNOMED CT to be used, the direction of the map must be from the non-SNOMED CT codes to SNOMED CT codes. A map designed to move data from code system A to code system B will serve poorly (if at all) 'in reverse' if it is used to map from B to A, unless all the links are exact semantic matches.

For analytics purposes where patient safety or data accuracy is important (e.g. point of care clinical decision support or data integration), it is important that the correlation of the map is an 'exact match' (or equivalence). For other purposes (e.g. epidemiology or care service delivery planning) it may be acceptable for the SNOMED CT code to be broader than (or a supertype of) the non-SNOMED CT code. However, broad-to-narrow and narrow-to-broad maps need to be used with care.

When a non-SNOMED CT code is being mapped into SNOMED CT, and an equivalent precoordinated SNOMED CT concept does not exist, a number of options are possible, including:

1. Map the code to a broader (supertype) SNOMED CT concept

   1. For example, map "DX0162: arthritis of left knee" to "371081002 |arthritis of knee|" with correlation 'narrow to broad'

2. Map the code to a SNOMED CT postcoordinated expression

   1. For example, map "DX0162: arthritis of left knee" to "371081002 |arthritis of knee| : 272741003 |laterality| = 7771000 |left|" with correlation 'exact match'

Designing and authoring maps requires expertise and appropriate resources. Large maps (e.g. tens of thousands of codes) are typically created and maintained by SNOMED International, National Release Centers, large healthcare organizations, specialist data suppliers and large system vendors. However, smaller maps may be created and maintained by smaller system suppliers, hospitals or clinics. Maps must be maintained to ensure that both the SNOMED CT content and non-SNOMED CT content remains current whenever either code system is updated.

A typical scenario requiring mapping to SNOMED CT is shown in the figure below. In this example, two source systems (using ICD-9 and ICD-10 respectively) are being integrated into a data warehouse using SNOMED CT as the common 'reference terminology' for analysis. Once this mapping is done, the same analytic techniques as used on native SNOMED CT records may be applied (See SNOMED CT Analytic Techniques).

Maps are represented in SNOMED CT's RF2 using a Simple map reference set, a Complex map reference set, or an Extended map reference set (depending on what additional information is required to support the implementation of the map). Code mappings are then performed by matching each non-SNOMED CT code in a patient's record with the 'mapTarget' field of the corresponding row of the map reference set, and using the SNOMED CT code found in the 'referencedComponentId'.

The UK Terminology Centre's Data Migration Workbench demonstrates some advanced uses of data migration and mapping products published by the UKTC, including Read Code Version 2 and CTV3 maps to SNOMED CT. A number of vendor products also map non-SNOMED CT codes to SNOMED CT for use in analytics, including Allscript's terminology service, Apelon's Distributed Terminology System, the Cerner Millennium Terminology (CMT) package, and Epic's electronic patient record systems.