Modeling Patient Safety Incidents Knowledge with the Categorial Structure Method

Abstract

Following the WHO initiative named World Alliance for Patient Safety (PS) launched in 2004 a conceptual framework developed by PS national reporting experts has summarized the knowledge available. As a second step, the Department of Public Health of the University of Saint Etienne team elaborated a Categorial Structure (a semi formal structure not related to an upper level ontology) identifying the elements of the semantic structure underpinning the broad concepts contained in the framework for patient safety. This knowledge engineering method has been developed to enable modeling patient safety information as a prerequisite for subsequent full ontology development. The present article describes the semantic dissection of the concepts, the elicitation of the ontology requirements and the domain constraints of the conceptual framework. This ontology includes 134 concepts and 25 distinct relations and will serve as basis for an Information Model for Patient Safety.

Introduction

Modeling patient safety information is essential for comparing and trending patient safety incident data across disciplines, organizations and countries; investigating and determining patient safety incidents and issues; examining the roles of system and human factors; determining the applications and limitations of existing strategies to reduce risk; and developing priorities and safety solutions [1]. This modeling is a prerequisite to operationalize and standardize reporting systems.

While patient safety information has been present in health care services for a long time, it is only in recent years that its growing development has enabled it to retain enough attention and allow an important increase of publications about patient safety. The need has therefore appeared to have a single comprehensive terminology resource, to offer researchers and managers a way to understand each other’s work and facilitate the systematic collection and analysis of relevant information from all available sources. The launch by the World Health Organization (WHO) of the International Alliance for Patient Safety (PS) has provided the opportunity to meet this need.

In 2009, the WHO Department of Patient Safety published the first report on a conceptual basis for an International Classification for Patient Safety (ICPS) [2, 3]. The report contained a list of terms and definitions of patient safety incidents and their circumstances. These terms were based on an analysis of the existing taxonomic classifications and definitions. A conceptual framework was provided as a basis for organizing the components in large notional categories comprising “key concepts” described by consensus definitions.

In its current form, the proposed conceptual framework aims to provide a consensual understanding of patient safety concepts but is not suitable for computer modeling. The main limitations lie in the structure of information: many relationships between concepts are not explicit enough to allow implementation in health information systems. An ontological representation including minimum requirements is needed, identifying the basic categories and their authorized and minimal relations, (i.e. the common elements in all patient safety domains: medication, clinical procedures, falls, medical devices, etc.) underpinning the standard information model needed to monitor and compare the patient safety issues.

The CEN (Comité Européen de Normalisation) Categorial Structure was defined as a minimal set of health care domain constraints to represent a biomedical terminology in a precise health care domain with a precise goal to communicate safely. It is a model of knowledge restricted to 4 constraints: a list of semantic categories; the goal of the categorial structure; the list of semantic links between semantic categories authorized with their associated semantic categories; and the minimal constraints allowing the generation and the validation of well formed terminological phrases. We already presented the merits of using the CEN/ISO standard on categorial structure to bind, harmonize and map between terminologies and ontologies [4, 5].

This article summarizes the PS-CAST project to propose a categorial structure for modeling Patient Safety information (PS-CAST is an acronym for Patient Safety CAtegorial Structure). The objective is to “dissect” and explicitly line up the elements of the semantic structure underpinning the broad concepts contained in the conceptual framework for patient safety, as a prerequisite for subsequent ontology development.

Background

Rossi-Mori has defined a three-generation classification of terminological systems: (1) traditional paper-based systems (first generation); (2) compositional systems built according to a categorical structure and a cross-thesaurus (second generation) and (3) formal models (third generation) [6]. WHO decided to initiate the development of its new terminologies under the second generation model as a starting point towards the third one. Current International Classification of Diseases (ICD - 10th revision) is an example of a first generation system. Ongoing development of ICD-11 aims to represent the relationship between concepts using a description logic (DL) like language [7]. PS-CAST is a contribution to this methodology within an application field of Patient Safety.

Issues surrounding the construction of an ontology about patient safety are common in several ongoing European projects such as Debug-IT (Detecting and Eliminating Bacteria UsinG Information Technology) [8], PSIP (Patient Safety through Intelligent Procedures in Medication) [9] or ReMINE with its RAPS (Risks Against Patient Safety) ontology [10]. Furthermore, in the past years, some conceptual information models were made:

• CDC’s Public Health Conceptual Data Model made by the Centers for Disease Control and Prevention (CDC) published in 2000. The purpose of the model was to document the information needs of public health and establish data standards [11].

• JCAHO Patient Safety Event Taxonomy, a common terminology and classification schema for collecting and organizing patient safety data in 2003 [12].

A drafting group of international experts formed under the auspices of the World Alliance developed a conceptual framework for ICPS over three years, using parts of JCAHO and other existing information models. A convergence was mandatory to ensure that the model would cover the main issues related to patient safety. A total of 48 key concepts have been defined to facilitate understanding. Each concept was tested for cultural and linguistic appropriateness using domain expert’s knowledge from several countries. The validity of the conceptual framework was also evaluated through a survey [13] and an analysis by technical experts on reporting and classifying incidents. The drafting group proposed 10 high level classes, as shown in Figure 1.

Figure 1.

The 10 high-level classes of the conceptual framework and some of their relationships.

These 10 high-level classes are separated in three groups:

• Clinically meaningful, recognizable categories for incident identification and retrieval (Incident Type, Patient Outcomes)

• Classes related to proactive and reactive risk assessment (Actions Taken to Reduce Risk, Detection, Mitigating Factors, Ameliorating Actions)

• Descriptive information (Patient Characteristics, Incident Characteristics, Contributing Factors/Hazards, Organizational Outcomes)

Also, two kinds of relations are proposed: Influences and Informs, that are related to the information flow (the white arrows).

To create a formal ontological representation of the Conceptual Framework, a mixed approach was decided. On one hand, the approach supported by the Artificial Intelligence Laboratory Universidad Polytécnica de Madrid (UPM) consisting in the semantic and ontology analysis of the reporting of the different incidents, using two already established examples (‘pressure ulcers’ and ‘falls’) from the Australian Patient Safety Foundation (APSF) and World Alliance for Patient Safety. On the other hand, our team worked on another approach, which is based on the semantic analysis of the conceptual framework followed by a PS domain-level ontology representation. Merging both approaches shall allow obtaining a complete domain ontology.

Methods

This work includes a development of a DL like or semi-formal ontology that adequately conceptualizes and formalizes the current conceptual framework. As requirement specification, this categorial structure has to be in relation with WHO classifications, considering: representational requirements for content (domain ontology), language instantiations (ontological lexicalization), and a desirable interoperability with SNOMED-CT, HL7, etc. The current ICPS documents were used as the main knowledge source, and work was closely done with domain experts from WHO, national agencies, etc., to derive any other necessary requirements.

The development of the categorial structure was carried out iteratively from the model proposed in the conceptual framework. The three main steps were: 1) to add one by one all the key concepts present in the conceptual framework. 2) to change the hierarchy of the concept categories according to elements added in step 1. 3) to add additional concepts to the conceptual framework to improve consistency and improve interoperability with international classifications, e.g. ICD-11, ICECI (International Classification of External Causes of Injury), ICHI (International Classification of Health Interventions), etc. and existing ontologies. At each step of this iterative development, the result was evaluated and validated first by one of the author and later on by the UPM team, the APSF team and Samson Tu from Stanford University.

There have been 17 successive versions with major changes. Each release was joined by a list of changes made in the ontology (add, modify, delete, move, etc.).

Here is an extract of the changes from version 09 to 10, submitted to the reviewer on March 17, 2010:

Proposition	Reviewer’s answer
Creation of concept : Harm_Type	OK
Move children of Patient_Harm in Harm_Type	OK
Creation of property : Has_Harm_Type (Harm_Type)	OK
Deletion of No_Harm_Incident, previously child of Incident	OK, this is a value and not a concept.

The development was also made by successive questions to the reviewer, for example on the version 12 on March 30, 2010:

Q : Health_Intervention concept is missing and have to be added to domain concepts. I propose to give it Characteristics properties and a target: Patient, is it consistent?

A : Health_Intervention has certainly Action_Characteristics and Action has target Factors or Patient or Health Professional. Also Health_Intervention and has means (associative relationship) Device and Process.

Our work can be described in 3 main steps:

1. Preliminary work

   1. Review of the definitions of key concepts.

   2. Semantic dissection of the conceptual framework: selection of the relevant key concepts.

2. Ontology development

   1. Representation of the concept categories with their granular concepts hierarchical relations and their associated relations.

   2. Testing the coherence of semantic relationships of the conceptual framework using the ontology representation and a reasoner.

3. Interoperability

   1. Alignment between Categorial Structure and ‘Falls’ and ‘Pressure Ulcers’ ontologies.

   2. Alignment between Categorial Structure and UPM ontology.

To obtain a full formalization in standard knowledge representation languages like OWL2 [14], we used an ontological conceptualization and formalization tool: Protégé [15]. The consistency of an ontology can be checked by reasoner that ensures that the categorial structure does not imply any incompatible inferences (classes can have some conflicting restrictions, especially when a class inherits from its parent properties). It also checks the structure itself, and it computes and analyses the subclass relations between every named class and recreate a complete class structure. This can be used to ensure coherence of what is understood by a computer when browsing the structural tree.

We recovered two examples of representation about ‘Falls’ and ‘Pressure Ulcer’ incidents created by the APSF [16]. More granular, they are the leaves of the domain ontology and we must ensure that our categorial structure is sufficiently substantiated to be linked. During the development of PS-CAST, we had to take into account the representations produced simultaneously by UPM [17], and we tried to reconcile them with our categorial structure. Proposals for changes have been made by both UPM and us to get a better understanding of each of the terms used and their justification [18].

Results

1). Preliminary work

The categorial structure complies with the key concepts of the conceptual framework. We have incorporated most of the key concepts except those pointing to meta linguistic entities that could find no place in the structure, for example, ‘class’, ‘concept’ or ‘classification’. We also estimated that some key concepts, for example ‘accountable’ or ‘resilience’, were not essential elements of a user interface for collecting incident (but could be useful in the analysis or evaluation of these events) and are therefore not included. We kept 39 out of 48 concepts in the categorial structure.

The definitions of the conceptual framework introduce fundamental notions of patient safety. We focused on 3 main categories:

• Patient Incident (Figure 2): specified by Incident Types (about 15 incident types, such as ‘Medication/IV Fluids’, ‘Blood/Blood Products’, ‘Nutrition’, ‘Oxygen/Gas/Vapour’) and Incident Characteristics. An incident is perceived by Detection, triggered by Hazards, and its effects can be modified by Contributing or Mitigating Factors.

• Patient and Organizational Outcomes: Patient related outcomes are the harm to the patient resulting from the incident and the Organizational Outcomes are the impacts upon the organization where the incident occurs.

• Actions Taken to Reduce Risk: preventive actions such as ‘Risk Assessment’, ‘Root cause analysis’… They are separated from Ameliorating Actions that are taken to compensate the harm after an incident.

Figure 2.

Representation of the Incident concept and its properties.

2). Ontology development

These three main categories are related to other classes by semantic relations. In the conceptual framework, these relations were not always clearly specified (by their definition) as being relational (has properties), hierarchic (is a) or mereologic (has part). Nevertheless it has been possible, in our ontology, to relate the majority of the classes to at least one of these main categories. E.g, according to Figure 2, Incident has a causal relationship with Hazards. Few classes has no relation with one of the three main classes. For example, the concept of Health is used within the definition of Health Care but by definition, a patient suffering from an incident is no longer healthy. Therefore, we have not implemented the relationship between this concept and others from the categorial structure.

The current version of the ontology includes 134 concepts, 25 distinct relations between them. After completing the categorial structure we were able to link elements of the conceptual frameworks to concepts from three different high level ontologies (BFO [19], Sumo or Dolce).

We checked the categorial structure consistency with Protégé integrated reasoner Fact++; no contradictory facts or errors were found, the inferred hierarchy was established logically valid.

3). Interoperability

We introduced many additional classes to the 39 key concepts in our categorial structure. We took the opportunity to improve interoperability with WHO classifications, such as concepts of ‘Managing Action’ and ‘Preventing Action’ from the ICHI categorial structure.

The categorial structure allowed to easily map 22 out of 25 (84%) main classes of APSF’s ‘Falls’ representation. The same alignment ratio could be found with ‘Pressure Ulcers’ incidents main classes. This mapping encouraged us to think that the categorial structure consistency allows more granular terminologies to be inserted. Table 1 presents an excerpt of an alignment between ‘Falls’ main classes and our categorial structure.

Table 1.

Excerpt of the proposed alignment between ‘Falls’ ontology and our categorial structure.

Falls Ontology Classes	CS Equivalent Classes
Action involved The type of action that the patient was performing at the time of the fall.	Incident_Characteristics > Origin_of_Incident
Documented at risk of fall Identify if the patient was documented to be at risk of a fall.	<no equivalent class found>
Restraint or safety devices failure and how or why it failed Select if there was a failure with a restraint or safety device	Contributing_Factor > Work_Environment
Fall Height Height of the fall	Incident_Characteristics > Reporting_of_Incident
History of falls Select to identify if the patient had a history of falls in the last 12 months.	Patient_Contributing_Factor
Mechanism of the fall Select to identify the mechanism of the fall.	Incident_Characteristics > Reporting_of_Incident
Known Falls Risk Prior to Falling Identify if prior to the fall the patient was assessed to be at risk for a fall.	Patient_Contributing_Factor
Floor Surface Identify the type of floor surface involved in the fall	Hazard > Physical Environment
Length of time the subject had been in the healthcare facility Identify the length of time the patient had been in the healthcare facility.	<no equivalent class found>
Medication Contribution Identify if medication use was considered to have contributed to the fall.	Healthcare_Services > Drug_Prescription
Object subject fell from Identify the object involved in the fall.	Incident_Characteristics > Reporting_of_Incident
Previous Fall this Admission Identify if the patient had fallen previously during the same admission.	Patient_Contributing_Factor
Actual Assistance or Supervision Ordered or in Use at Time of Fall Select if assistance or supervision had been requested or ordered.	<no equivalent class found>

We sent to UPM intermediate versions of our work and parts of the report concerning it. UPM managed to align their representations with the categorial structure while keeping their already created value sets. In the final version, class names produced by UPM are very similar to those of the categorial structure and the connection with our ontology is quite straightforward. This interoperability ensures the validity of the categorial structure as a domain ontology to which all current and future incidents developments, such as ‘Falls’ and ‘Pressure Ulcers’, can be linked easily.

Discussion

Summary

PS-CAST adds to the conceptual framework properties suitable to elicit an organization of concepts in a normalized hierarchy, according to the classical relation of subsumption. The categorial structure we present allows to organize elements of the PS domain within a well built formal representation of patient safety knowledge with concepts and relationships able to integrate the more granular value sets proposed by UPM and the World Alliance for Patient Safety. Finally the categorial structure can be linked to high level ontologies.

Related work

The ReMINE project [20,21], in the deliverable D4.2 [21], propose an ontological representation for managing risks against patient safety (RAPS). Although the structural organization of this ontology is different from the conceptual framework, and ICPS interoperability was not an objective of the authors. Also, the merging of multiple ontologies (such as BFO) lead to adding many non domain-specific concepts. We note that the authors have kept some key concepts in their conceptual model but some concepts are missing such as ‘mitigating factors’, ‘preventing actions’ or ‘detection’. Furthermore, D4.2 provides an appendix with a list of concept definitions in natural language that allow for immediate processing into formal language.

At the Medical Informatics Europe 2009 conference in Sarajevo, Schulz et al. made an appraisal of ICPS [23] and concluded that ICPS is neither a classification nor a taxonomy, but a general conceptual framework that presents properties necessary for modeling as an ontology. The categorial structure we present in this document addresses issues related to formalization and described by Schulz et al. for the conceptual framework.

In a recent article [22], Ceusters announces the creation of an adverse event ontology. It introduces 9 concepts missing from deliverable 4.2 of ReMINE [21] such as harm, mitigation or prevention and which are present in our Categorial Structure. He however emits criticisms about some ambiguous definitions within ICPS such as “class” or “semantic relationship”. We have not included these concepts in our Categorial Structure for the same reasons, and because they are not domain-specific. In a second argument, Ceusters explains that “ICPS further considers the terms ‘event’, ‘circumstance’, ‘situation’ and ‘factor’ all to be more generic than ‘incident’”. We completely agree as in our ontology, Incident is now a child of Event. Furthermore, to remedy these problems, Ceusters indicates that additional efforts must be provided, using an ontological methodology; we believe that we have tied to achieve these efforts. Concluding his article, Ceusters states that the alignment of the adverse event ontology (under development) with ICPS would be necessary as ICPS “enjoys a broad domain coverage” despite a “lack of formal rigor”. We reviewed all key concepts definitions within ICPS and altered the relationship between these concepts if necessary, we formally produced a more rigorous and consistent model, with good practices in ontology building.

Perspectives

It is desirable that other domains than ‘Falls’ and ‘Pressure Ulcer’ are also documented in order to answer other questions related to patient safety (e.g. nosocomial infections and safety incidents occurring during surgery). These terminologies, whose concepts are more granular, have to be mapped with concepts of the PS domain, for homogeneity purposes. It is also recommended to bind the other terminologies in the most appropriate place. For example, ICECI may be related to the concept Incident Characteristics and ICD-11 found its place in both Patient Characteristics (Reason for Encounter) and Harm. We considered several ways to attach the WHO classifications in the ICPS ontological structure (e.g. in Patient Outcomes or in Patient Characteristics) and we are now investigating the appropriateness and compatibility of the respective ontological dissections. The queries made on the database may be based on value sets collected in terminologies and thus express all types of incidents for a given disease.

PS-CAST is a contribution to the new WHO methodology for the development, maintenance and translation of international terminology resources. This methodology is based on collaborative work between domain experts and knowledge engineers. WHO is currently experimenting this approach for the ICD-11 revision. Our Categorial Structure aims to identify the basic categories and their authorized and minimal relations as the ontology framework underpinning the standard information model needed to monitor and compare the patient safety issues locally, nationally or internationally. The final output of this work cannot be a first generation terminological artifact like the ICD-10, but a new generation system, like ICD-11. This development can lead to new terminological artifacts based on an information model adapted to the digital revolution in recording and processing health care data.

Table 2.

Comparison of ReMINE and PS-CAST.

	PS-CAST	ReMINE
Objectives	Create a Categorial Structure	Create a realism-based ontology.
High level ontologies	BFO, Sumo, Dolce (a posteriori)	BFO (a priori)
Constraints	Minimal	Maximal
ICPS Conceptual Framework conformity	Complete (nearly)	Partial
Final users	WHO’ institutions partners, familiar with Patient Safety	Independent users, familiar with ontology modeling