Health information model

From Endeavour Knowledge Base

This article describes the approach taken to producing information models, including ; what they are, what their purpose is, and what the technical components of the models are.

The article does not include the content of any particular model.

What is a health information model?

In this context, a model is considered as a representation of the arrangement of data within health records, together with a means of representing Query in a way that a human can understand and a machine can interpret.

There are as many models as there are business processes in healthcare. It is impractical to consider a single 'standard' model of healthcare data. Apart from the fact that the nature of health data is constantly changing, there is rarely agreement as to what the data should be, beyond that needed for a particular set of business processes.

The lack of a single health and care common model, or 'standard' leads to the challenge of how systems can interoperate with each other in the absence of a standard model, without loss of semantic meaning.

This challenge is addressed by the Discovery information model.

By convention, the approach to this problem has been to establish specific basic models with agreed structure and content, and create extensions or variations on them for specific purposes. Examples of this approach are HL7 FHIR and OpenEHR. Such agreed models have enabled systems to communicate with each other by sharing or exchanging data, and have improved the efficiency of many health management processes resulting in better health outcomes.

The approach taken to modelling in Discovery is a little different but consistent with the above. The approach recognises that semantic interoperability can be achieved by the use of a shared language , leaving specialised modelling for particular business processes. In other words, just as humans understand each other by sharing a grammar and vocabulary and a protocol such as written text, so can machines.

This is not a new idea. This is the basis of the semantic web and the set of languages that have developed to support it.

Unlike human language, machine based languages must use logical constructs because computers can only operate using one of the basic logic gates. Humans can say illogical things such as "it was the best of times, it was the worst of times". Systems that comply with the agreed grammar cannot, or at least, should not.

There is still a need to agree models for particular businesses but as these use the same grammatical rules and use the same vocabulary, it is much quicker to create them. If system A sends data in model M1 and system B receives that data, but the users of system B do not use model M1, then system B can still accept and interpret the data from system A without loss of meaning.

The approach eliminates the need for "standardised" schemas. Instead, schemas are defined by the languages. Interpreters are needed to translate the words if a system uses its-own dialect, but those interpreters operate using a semantic web standard language.

The remainder of this article considers how models and ontologies can be constructed using this approach.

Visualisation

Types of data as a graph


The data in a health record stored can be visualised as a graph, and a model of health data can be visualised as a graph of types of data,

In a graph, a node can link to many other nodes and links are constantly changing , similar to the way the human brain appears to operate. This intuitive approach means that there are as many potential nodes and relationships (vertices and edges) as needed and they can change rapidly.

This contrasts the approach with the relatively fixed structures often adopted in healthcare Systems.

The example on the right is entirely arbitrary but illustrates a problem. What does "condition record" mean, or indeed what is a "condition"? Why is a patient linked to a person and what does "linked to" mean?

The answer is that the "terms" or "concepts" used in a model should be derived from a vocabulary whose terms have meaning and are formally defined. Some terms have meaning in whatever context they are used whereas others have different meanings in different contexts. In defining terms, it is necessary to defined them precisely enough for a computer to interpret the meaning safely i.e. the context of an idea is part of the idea itself.

The most difficult challenge is to agree the definition and meaning of the concepts in the context they are used. The agreement as to a particular model is less important. A definition defines a concept in relation to other concepts. Within a domain of interest such as healthcare, all concepts are indirectly related in some way to all other concepts in that domain.

Luckily, standards have evolved to enable machine readable definitions.

The crucial step in the discovery approach is to apply this principle to both the things that are being recorded (such as clinical concepts), as well as the structure of entries in records themselves.

Semantic Web

The semantic web approach is adopted. In this approach, data can be described via the use of a plain language grammar consisting of a subject, a predicate, and an object. A triple. The theory is that all health data can be described in this way (with predicates being extended to include functions).

The consequence of this approach is that web standards can be used such as the use of Resource Descriptor Framework or RDF. This sees the world as a set of triples (subject/ predicate/ object) with some things named and somethings anonymous. Systems that adopt this approach can exchange data in a way that the semantics can be preserved. Whilst RDF is an incredibly arcane language at a machine level, the things it can describe can be very intuitive when represented visually.

Put together with graph this means that a graph can be organised with subjects as nodes, objects as nodes, and predicates as relationships.

In other words the Information modelling approach involves an RDF Graph.

Concepts, versus codes

A general language consists of a vocabulary of words arranged according to a syntax that follows grammatical rules.

Information consists of ideas. Another word for ideas is a 'concept' . A concept may be named, in which case the meaning of the concept can usually be understood by the name, or they may be unnamed (an "expression") which is a definition made up from other named or unnamed concepts. The term "chest pain" implying the idea of a pain in the chest is one example of a named concepts. "Cheat pain, worsened by exercise" is an example of an expression style concept made up from the concept of "chest pain", the statement that it is "made worse by" and the statement that it was made worse by "exercise".

The new generation of health record management systems tend towards the recording of concepts with an objective to closely match the idea behind the entry. These types of concepts are often called term based concepts.

A modern term based concept is defined in relation to other concepts by a set of assertions indicating whether the concept is equivalent to, or a subtype of, a set of other concepts. The standard approach to this is via the use of Description Logic (DL). By using DL, a computer can automatically classify a concept. Snomed-CT is the worlds largest ontology of term based concepts.

Coded concepts, originate from a different starting point. The intention of a coded entry is to pre-classify an entry before it is recorded. The classification marker or code, is designed for a particular set of business processes e.g. analytics or payment. A coded concept, being pre-classified, relies on categorisation of the codes, which may or may not imply subtypes. Consequently, as the philosophy is different, code based concepts have to be dealt with differently from term based concepts.

Because of their history, it is not always possible to assert the exact meaning of a code based concept. However, it is often the case that meaning can be inferred or approximated from a coded entry. This inference is achieved via the use of maps.

A map is a sort of statement that something is possibly or probably similar to something else. It has much less weight than an asserted relationship. Code based concepts can be mapped to term based concepts which enable the use of the vast volumes of data already recorded in systems. Maps generally fall into 4 patterns. These are illustrated in the context of code based concepts as follows:

  • A coded concept has one map which is mapped to one term based concept with a certain weighting or map category. For example the icd10 code for Angina has a map which maps to a single term based Snomed-CT concept of angina, with a category indicating that the source concept is properly classified. Note that many coded concepts may be mapped to one single term based concept. The map is viewed from the perspective of the coded concept.
icd10:I209 |Angina Pectoris (ICD10 I20.9| 
            has map [mappedTo sn:194828000 |Angina (disorder);
            mapCategory sn:447637006 |Map source concept is properly classified]

A coded concept has more than one map and each map may map to ore than one potential term based concept


The approach taken in Discovery is to classify according to "sets" and thus adopting the approach taken by modern ontologies. A set is a definition of a set of things that have the same properties ( i.e. a class and a set are the same thing). Sets of ideas may contain other subsets which are objects that have the same and more specific properties than the super set (or super class) i.e. a subclass of a superclass.
Putting together RDF and sets, the net result aligns with RDF, RDFS and OWL2 i.e. the ontology web language. The vocabulary of OWL2 is used to precisely define concepts in relation to other concepts. OWL2 uses an underlying idea of "Description Logic" which is a way of defining things in a logical and consistent way so that a classification can be reliably produced.

Data sets and schemas

Having a grammar and a vocabulary represented in RDF and OWL is not enough. To model things for specific purposes it is necessary to describe precise structures. These may be referred to as data sets or schemas, or more commonly, information models.

N.B. IN Discovery, the term information model encompasses structures, semantics and query.

A data set takes rather vague general statements and arranges concepts in precise manner. This aligns precisely with the semantic web language Shape constraint language (SHACL). This differs from OWL in that OWL constructs assume a partial view of the world with everything not described being uncertain, whereas SHACL states how something should be.

To support both machine readability and standard based interoperability, Discovery adopts the necessary elements of SHACL in addition to RDF/ RDFS and OWL

Interrogating data for information.

Having established a representation of data using a set of grammars (RDF/RDFS/OWL/SHACL) it is necessary to represent a means by which the data could be interrogated to produce useful information.

Once again the semantic web community has established a machine readable common grammar for query known as SPARQL. SPARQL is designed to ask questions of RDF and is thus an ideal way of representing query logic.

Dialects and alternative languages

It is all very well supporting semantic web standards, but the world often adopts alternative approaches.

To that end, Discovery modelling tends to support grammars and syntaxes that are in common use, as long as they do not distract from the core models.

Such as examples include Expression constraint language (ECL) which is a way of expressing entailment queries of complex class expressions. Another includes GRAPHQL which is a way of querying constructs by presenting a template of expected results.

Information model APIs and languages

For an information model to be useable, it has to be accessible in some way. The means of accessing an information model is via the use of a language i.e. an information modelling language and this is described in a separate article. The language assumes a graph representation of the model and uses RDF concepts as its basis.

IM Service architecture

For an information model to be useful, it has to have at least one information model service, i.e. an operational service that provides access to one or more information models. A service must provide a set of APIs as well as provide instances of the model for implementations to use directly should they wish to.

The diagram on the right shows a tiered architecture for such a service. Information model APIs are described in a separate article.

All implementation code including the evolving service, APIs, language grammars and object models are also available on Github in the following repositories:

https://github.com/endeavourhealth-discovery/IMAPI

Utilities that use it and transform between syntaxes are at:

https://github.com/endeavourhealth-discovery/InformationManager

A viewer of the information model is at:

https://github.com/endeavourhealth-discovery/IMViewer

Information model purposes and functions

The information models have 4 core functional requirements internal to a model: Description of the model , validation of model content, population of the model, and query of the model. In support of query there is also the need to support inference which generates new insights that were not necessarily authored.

In addition the information model must support the same 4 core functional requirements on actual health data that is modelled.

Systems that use the models can use any or all of three approaches:

  1. Direct use of the model data content as a database (or set of files that can populate a database via script)
  2. Use via a set of APIs (both local and remote) designed to provide access to the data within the model, or to trigger outputs of the model for 1)
  3. Use of the information model technologies themselves via the use of the published open source code

The main functional purposes of an information model is further described:

  • Description of the model. There is little point in having a model unless it can be described and understood. Knowing what is in a model is a pre-requisite to using it. For example, there is no point in trying to find out if a patient record indicates whether or not they have diabetes if the model doesn't include the ability to record it. In order to understand a model, two techniques are required: diagrammatic representation and human readable text representation. A model must support both.
  • Data Validation is essential for consistent business operations. Data models, user input forms, and data set specifications are designed to enable data collections to be validated. Maintaining a standard for data collection is essential. For example, if you have a patient record in front of you, you will likely need to know their approximate age. To work this out date of birth must be recorded. Validating that the date of birth can be and has been recorded is important. However, if more than one date of birth was recorded for the same patient, it would be less valuable. Thus a modelling language must include the ability to constrain data models to suit particular business needs as part of validation, even when the data model shows more than one.
  • Population of the model. It is impractical to build model content from scratch and likewise virtually impossible to populate instances with existing data without some manipulation. An information model must contain the ability to model mappings between currently held data and model conformant data.
  • Enquiry (or query) is necessary to generate information from data. There is little point in recording data unless it can be interrogated and the results of the interrogation acted upon. Thus a modelling language must include the ability to query the data as defined or described, including the use of inference rules to find data that was recorded in one context for use in another.
  • Inference is pivotal to decision making. For example, if you are about to prescribe a drug containing methicillin to a patient, and the patient has previously stated that they are allergic to penicillin, it is reasonable to infer that if they take the drug, an allergic reaction might ensue, and thus another drug is prescribed. Thus a modelling language must include the ability to infer things and classify things for safe decisions to be made.


Example of model content basic assumptions

In constructing a model of health data, it is necessary to have an agreement as to the sort of things that a model will contain and how they will be categorised.

It is fair to say that there will probable never be a universally accepted approach to this problem, but nevertheless, any information model needs to at least put a few markers down.

Healthcare modelling approaches such as hl7 and openEHR have each made some basic assumptions as to their respective starting categorisations. They are however incompatible and as a result, transfer of information between systems using the different approaches has proved expensive. The fall back position has been to continue with whatever model a particular system has and progress is delayed.

A safe starting point is to consider some categorical terms that are unlikely to be controversial and would be consistent with the open standards in place. For the sake of making a start, the following categorisations are proposed: Event, Entry, Provenance, ontology, types, state, query

  • Everything that is recorded starts with an event. In this context an event is a machine level event that signals a change of state or a desire to change a state. The event is usually associated with a description of what the event is and some data associated with the event. The data associated with the event normally includes the intention, such as a desire to add/amend or delete data in a record, as well as the data which was recorded as part of the event.
  • The net result of an event is the creation/update/deletion of, an Entry in a health record. The term ‘Entry’ is used in its intuitive meaning here. If one were to look at a record it would consist of entries, not events.
  • Because an entry is generated from one or more events, an entry has provenance. Provenance enables the audit and validation of an entry, including all events that led to the state of the current entry. A subset of an entries provenance is the “audit trail”, which is pivotal for medico legal purposes.
  • An entry in a record has a number of attributes which describe the entry. For an information model to succeed there must be an agreement as to what these attributes mean. This is achieved by the use of a shared Ontology. An ontology precisely defines the meaning of an attribute, and the type of values that an attribute might have. This means that ANY data can be exchanged as long as an entry uses attributes from the agreed ontology.
  • Agreement on the definition of concepts is not enough. Agreement on context is also important. Most would agree that a date of birth is the date a person on was born. But what about an entry in a record for Diabetes? Does it mean the person has the condition or does it mean the clinician is considering the condition? Context is provided by the ontology also but must use an ontology structure that can preserve context.
  • There are a huge number of business processes in healthcare. Each business process is associated with a requirement to exchange data that is relevant to the business. This is partly achieve by assigning types to entries. Types indicate the main purpose of the entry. An agreement as to what the types are, and consequently, what the associated attributes of an entry of a type should be, and what the values of the attributes should be, is essential for business.
  • It is generally the case that an entry can be considered as either representing an event in time (a different use of the word event) or a persistent state. Technically these categories are conceptual rather than real but are important for business level modelling. For example, a date of birth might be considered as a state and therefore might be modelled as a cardinal of 1 against a person, even though a series of historical entries have recorded a date of birth. State can be described by the use of types to indicate state versus event entries to indicate things that happened but do not persist. Many types are both.
  • Put together this equates to an ontology of concepts which are used as types, attributes and values, together with structural definitions of their relationships for context and business purpose. Terms used to describe these things are purely convention ; resources, resource profiles, archetypes, templates, value sets, dataset definitions are all simply ontological relationships.
  • All of this is irrelevant unless entries can be queried. Query itself produces new structures such as the above. Consequently a means of querying a records, which are projected as a graph is needed.

Model structure and content

Surprisingly, with the use of an agreed ontology and an agreed way of representing it via an open standard language such as the information modelling language, there is no real need to have one model structure.

Content of a model, including the definition of types, is driven entirely by the business which it is designed to support. A specialist in immunology is likely to need different content than a General Practitioner. However, there needs to be an agreement on what the concepts in use mean, particularly in context. Otherwise data cannot be exchanged.

The information modelling language means that one can have as many information model instances as needed. The language is like any other language but with some logical constraints. It may be possible to model the novel of War and Peace, but to state that "it was the best of times, it was the worst of times" is NOT allowed.

Thus the common information model is in fact no more than a model that models information as used in a common way. The idea that somehow models can be "Standardised", is somewhat quaint unless the business itself is standardised. If the business is standardised (i.e. everyone agrees to do the same thing) then a common model is a standard.

Thus in the Endeavour Discovery model the only standardisation is:

  1. The basic assumptions as to the difference between events, entries, and their provenance
  2. The selection of the best fit ontologies for particular purposes, as long as those ontologies conform with the information model language constructs, which enable world wide adoption by the systems that already use the language

For the content of the models themselves this can be accessed through the IM viewer (under development) or by downloading the model and viewing via a generic RDF graph viewer.

The approach to modelling covers 3 aspects of health record information:

  1. Models of data stored in health records and their supporting records.
  2. Ways of retrieving data both from the model itself and the health record data stored, i.e. various forms of query.
  3. Models of maps between originally entered data and a selected model designed so that one semantically defined query will pick up data entered in a variety of ways.

Ideally a model should be designed both for human visualisation and for computers to use. This is the approach taken to the Discovery information model.

This article describes the meta data model of an information model (and does not include the content of a particular model. The article makes reference to the languages that may be used to access the model, using either interoperability standards or a pragmatic approach, and this language is described in the article introducing the health modelling language.

The information model component types can be illustrated as follows:

Main information model component types