gms | German Medical Science

49. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie (gmds)
19. Jahrestagung der Schweizerischen Gesellschaft für Medizinische Informatik (SGMI)
Jahrestagung 2004 des Arbeitskreises Medizinische Informatik (ÖAKMI)

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie
Schweizerische Gesellschaft für Medizinische Informatik (SGMI)

26. bis 30.09.2004, Innsbruck/Tirol

Use of a new technology and Markov models to record and analyze the triage procedure in an emergency centre

Meeting Abstract (gmds2004)

Suche in Medline nach

  • corresponding author presenting/speaker Christelle Despont - Hôpitaux Universitaire de Genève, Genf, Schweiz
  • Olivier Rutschmann - Hôpitaux Universitaire de Genève, Genf, Schweiz
  • Josette Simon - Hôpitaux Universitaire de Genève, Genf, Schweiz
  • Christian Lovis - Hôpitaux Universitaire de Genève, Genf, Schweiz

Kooperative Versorgung - Vernetzte Forschung - Ubiquitäre Information. 49. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie (gmds), 19. Jahrestagung der Schweizerischen Gesellschaft für Medizinische Informatik (SGMI) und Jahrestagung 2004 des Arbeitskreises Medizinische Informatik (ÖAKMI) der Österreichischen Computer Gesellschaft (OCG) und der Österreichischen Gesellschaft für Biomedizinische Technik (ÖGBMT). Innsbruck, 26.-30.09.2004. Düsseldorf, Köln: German Medical Science; 2004. Doc04gmds076

Die elektronische Version dieses Artikels ist vollständig und ist verfügbar unter:

Veröffentlicht: 14. September 2004

© 2004 Despont et al.
Dieser Artikel ist ein Open Access-Artikel und steht unter den Creative Commons Lizenzbedingungen ( Er darf vervielfältigt, verbreitet und öffentlich zugänglich gemacht werden, vorausgesetzt dass Autor und Quelle genannt werden.




This last thirty years, emergency centers (ER) have known an increasing number of consultations [1]. Therefore, health proficiencies have searched measures to improve the flows. One proposed answer is the set up of a triage process to state care priority based on emergency level and seriousness of cases [2].

The triage function is assumed by experimented nurses. The triage is an evaluating of the emergency level based on a short collection of information about the consultation's motive, constants and observations [2].

To perform this task, triage nurses need a working tool. In the University Hospital of Geneva (HUG), this tool is a paper based two sided page. The front side contains fields about anamnesis and the back side a list of motive codes for consultation with available emergency levels. This paper is used as a transmission support and as a working paper to state about the main problem presented by the patient. The lifecycle of the paper ends when the patient is taken for active care. Between the choice of an orientation and the active care, the triage form could be completed with observations, new constants, etc.

The tool has been constructed to answer staff's needs, based on other similar researches and their investigations. Since its introduction in 1997, the triage form has evolved. However, the concrete use of this tool has never been systematically described. Such kind of processes in this context is really difficult to observe and record.

During the trial of a new technology in the ER, we have proposed to record components of the triage procedure. The digital pen technology is a real innovative tool which opens new perspectives for all semi-structured or structured handwritten data acquisitions. This technology allows nearly real-time digitalization of texts written on paper based forms. In addition, the pen stores information as timing and succession of fields included in the process.


During writing process, the digital pen records dynamic information (movements, timing, etc.) using a camera and a pattern printed on the form with the standard layout of the document for localization. When the pen is connected to computer, data is transmitted. An Application Service Handler (ASH) deployed on the server gets information and treats it to store under different format. For ER center 3 formats were chosen:

1. High quality JPEG pictures of the whole forms

2. SVG files on which a batch afterward adds animation skills.

3. Storage of information about all written strokes in a Oracle database including for each stroke:

a) Unique identification of the pattern (also used for identifying pictures)

b) Absolute stroke start time in [ms] and stroke end time

c) Identification of the field of the form in which the stroke has been written.

At the end of the trial period in ER the treatments of stored information started. All processing have been made using JAVA to access and treat pictures content and database records.

Batch treatments concern the three stored formats. JPEG pictures representing strokes written by users have been merged with the layout of the form. This treatment has been performed using JAI library to simplify lecture of content for physician and nurses. All SVG files have been treated to add animation skills. Such adding is useful to visualize a specific process and to access to previous information when data is corrected. Finally database was accessed to extract information about timing and components of the process in terms of fields completed.

Among all questions asked by professionals, the first is to extract, if existing, the generic path followed by nurses through the form to perform a triage. Such question requires the extraction for all forms of the completed field sequences and to find a meaningful representation to synthesize this data.

Two approaches have been used. The first one is histograms of accessed fields at different step. A new "step" starts when users access to a field which is different from the previous one. There is no direct information about the transitions but if histograms show few dominant fields at each step, a path tendency can be built.

The second approach uses 1-order Markov models for describing evolution of processes over time. At every instant t the model M is in one state. The quantity pAB, indicated on each transition between states, denotes the conditional probability that at step t+1 M will be in state B given it was observed in state A at step t [3]. We do not know what is more meaningful for medical practice between considering past or future transitions. The two models have been sketched: a graph with output transitions and one with input transitions. Input transitions have nearly the same meaning than output transitions described before: pAB denotes the conditional probability that at step t-1 M was in state A given it is observed in state B at step t. Each state of the model is a field, or a combination of fields in the triage form.

After description of the problem with complete Markov models some grouping together appears more eloquent. Health professionals have defined meaningful grouping and a second processing has been performed to extract new graphs. Multiple criteria can be added to explore different aspects of the problem.


In this paper are presented technical results and extractions obtained without medical interpretation of the processes.

During the seven days trial of the technology in ER, 1183 entries were recorded by IAO. Among them, 954 forms have been digitalized with a completeness rate of 87%. To limit the error in the extracted sequences, only forms containing a motive of admission, an emergency level and an orientation have been used to build Markov models. These fields are considered by professionals as the triage conclusion and 907 forms reach this criterion.

The histograms of different steps are meaningful for the three first steps, but afterward distribution is quite fair between fields. There is no unique path through the forms showing that individual practices have more weight. For this reason the preference was given for the Markov models approach.

The first graphs sketched were as complete as information available has allowed. Because of no teaching process about how to use the pen, triage nurses have used it as it was a standard pen. We observed a lot of writing out of bounds (approximately 20% of forms concerned) and strokes crossing multiple fields. Hopefully, strokes covering multiple fields are recorded twice in database with the same start time and the same end time. Thus a possible criterion to avoid false new step is to remove these strokes during extraction of sequences. Despite this treatment, some fields knew a lot of oscillations between them. It appears that these fields are generally not clearly distinguished in clinical practice. Because we do not work with content and in accord with medical staff, grouping together has been made to avoid oscillation transitions and to make the graphs more readable and interpretable. From the 56 initial fields (30 are constants and 11 for pain scale) we keep nine fields and add a start state and an end state.

The graphs presented in figure 1 [Fig. 1] concerns output transitions. The complete sequences have been uses to build transition matrix without stop criteria such as "orientation defined". It highlights that the triage process includes three main entry points (97% of all entry transitions), with one dominant: General information about date, hour, and is the patient alone or not. However there is no consensus for end transitions. In addition there is no dominant path through the triage form and some states are not reached by the most important transitions (among the 75% indicated). Among transitions between states during the process, some of them encounter a high agreement between nurses like transition from admission motive code (Code Motif) to emergency level (Degrés urgence).

In addition, states in the model have been arranged as they appear on the triage form, showing spatial displacements on the triage forms. Such layout allows detection of design impairments of forms. It provides arguments for need of redesign or arguments to back up choices made.


This kind of model highlights the complexity of the triage procedure and create a lot of new questions among professionals like in the 75% of transitions none goes to "Constants" state, how are related constants and triage criterion, etc. As explained in the methods section a lot of grouping and filter criteria can be implemented to focus on these questions and provide clinician data to explore their procedures.

Among actual existing tools, the digital pen technology seems really natural and non invasive to record handwriting processes of data acquisition. Completed with observations, information collected provides a particularly good view of a complex procedure without heavy intervention.


Roux L. Infirmier d'accueil et d'orientation. In: L&C Es, editor. Urgence 2004; 2003; Paris; 2003. p. 587-593.
Divorne L. Démarche de tri: outils de tri existants. In: L&C Es, editor. Urgences 2004; 2003; Paris; 2003. p. 595-602.
Jarvis JP, Shier DR. Graph-Theoretic Analysis of Finite Markov Chains. In: (Eds.) TW, editor. Applied Mathematical Modeling: A Multidisciplinary Approach: CRC Press; 1999