Development and testing of a hierarchical method to code the reason for admission to intensive care units: the ICNARC Coding Method

J. D. Young1, C. Goldfrad2 and K. Rowan2

1Nuffield Department of Anaesthetics, Radcliffe Infirmary, Woodstock Road, Oxford OX2 6HE, UK. 2Intensive Care National Audit and Research Centre, Tavistock House, Tavistock Square, London WC1H 9HR, UK*Corresponding author

{dagger} On behalf of the ICNARC Coding Method Working Group.

Accepted for publication: May 25, 2001


    Abstract
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
A computer-based hierarchical method was developed to code conditions leading to admission to intensive care in the UK. The hierarchy had five tiers: surgical status, body system, anatomical site, physiological or pathological process and medical condition. The hierarchy was populated initially using the free-text descriptions of the reason for admission from 10 806 admissions recorded as part of the Intensive Care Society’s UK APACHE II study. After refinement and error-checking, a prospective evaluation was undertaken on 22 059 admissions to 62 UK intensive care units. Individual units coded between 60 and 1610 (mean 356) admissions. All but 50 (0.2%) of the admissions could be coded and 38 units coded every admission. Fifty admissions (0.2%) could not be coded within 24 h of admission but were coded subsequently when more information became available. Of the admissions, 96.1% were coded at all levels of the hierarchy in the coding method. Six hundred and thirty-seven of the 741 unique conditions (85.9%) were used in one of the five reasons for admission and 564 (76.1%) in the primary reason for admission. Five conditions account for 19.4% of all primary reasons for admission. This is the first method to be developed empirically for coding the reason for intensive care admission.

Br J Anaesth 2001; 87: 543–8

Keywords: records, medical, computerized; intensive care


    Introduction
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
A systematic method for recording the reason for admission to intensive care is central to the analysis, interpretation and comparison of mortality and morbidity data collected in different units, countries or at different times. The conditions have to be coded to ensure consistent terminology and to allow statistical analysis and modelling. The Intensive Care Society’s UK APACHE II study13 used a modified version of the American APACHE II ‘primary indication for ICU admission’ classification method to code the reason for admission to intensive care into multiple categories. In spite of the modifications, which removed impossible combinations of the seven body systems and 36 precipitating factors from the American APACHE II classification (e.g. the combination of body system, gastrointestinal with precipitating factor, and valvular heart disease), problems were still encountered by study participants trying to code the reason for admission.

Many of the problems arose because the categories ranged from discrete, well-defined conditions (e.g. diabetic ketoacidosis) to groups of conditions (e.g. peripheral vascular surgery) and were neither comprehensive nor exclusive. In addition, it was often difficult to identify a single primary reason for admission and there was no provision for coding other important conditions occurring either during or after the first 24 h. If the American APACHE II rules were followed strictly (only information available during the first 24 h to be used to code the reason for admission), there was no provision to update the reason for admission if it could be provided, either in more detail or more appropriately, after the first 24 h.

Moreover, the categories could not be cross-mapped to either of the two main diagnostic coding methods in use in the UK, the Read codes [4] and the International Classification of Diseases (ICD-9/10). The lack of precision in the categories also impaired the ability to compare the mortality rates for single conditions with published case series.

As part of its remit to establish a national comparative audit of patient outcome from intensive care, the Intensive Care National Audit & Research Centre (ICNARC) convened a working group to develop and test an improved method for coding the reason for admission to adult and paediatric intensive care and high-dependency units. The primary requirement for the method was that it would add explanatory power when attempting to estimate the probability of hospital death after intensive and high-dependency care. Other requirements included the ability to describe the workload of an intensive care unit (ICU) in order to be able to compare it with other units’ workloads and to relate a unit’s activity to the workload of the hospital. The coding method would have to be able to code multiple reasons for admission in sufficient detail to be of value to clinicians and be capable of cross-mapping to other coding methods. As the coding method was to be used in conjunction with physiological data, it had to describe the underlying cause of deranged physiology (e.g. splenic trauma) rather than simply duplicating physiological information by describing the pattern of physiological changes (e.g. hypovolaemic shock).

This paper describes the development, testing and final form of the coding method, now called the ICNARC Coding Method (ICM).


    Methods
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
Literature review
A literature search was undertaken to locate any published intensive care-specific coding methods that predated or postdated the APACHE II method. As none were found, direct approaches were made to groups known to be active in the field to identify any unpublished methods. The following unpublished methods were identified: the Fichier Diagnostique (January 1994 update), the classification method of the Société de Réanimation de Langue Français; a list of conditions produced collaboratively by the developers of the Mortality Probability Models (MPM II) and the Simplified Acute Physiology Score (SAPS II); a list of conditions forming the interim results of the Diagnosis Task Force from Project IMPACT in the USA; two other lists of conditions based on ICD-9 and Diagnostic-Related Groups; and the diagnostic coding methods used in the development and final implementation of the APACHE III method. From the systematic search of the published and unpublished literature, it was apparent that none of the existing methods for coding the reason for admission to intensive care had been derived empirically.

Outside intensive care, other generic diagnostic and procedural coding methods were available. The three in most common use in the UK were the ICD-9 or 10, the Read codes (or CTV3), developed by Dr James Read for computerized records in general practice, and the Office of Population and Censuses and Surveys procedural coding system (OPCS4). These were all exhaustively comprehensive, but this was also their disadvantage in that they contained large numbers of conditions that were irrelevant to both intensive care and high-dependency care and would, therefore, be very cumbersome to use.

Development of a coding method
A free-text description of the reason for admission had been collected for each of 10 806 admissions to the 26 ICUs that participated in the UK APACHE II study. In most cases, this free-text description was augmented with an APACHE II diagnostic code. This provided a core data set from which the coding method was developed.

Before analysis of the qualitative data, a theoretical structure for the coding method was proposed, based on the available published and unpublished methods, and consensus within the working group of seven senior intensive care physicians and researchers. A computer-based, hierarchical structure, coding by body system, then by anatomical site, then by physiological or pathological process (‘system: site:process’ structure) was proposed. A prototype hierarchy was prepared for the respiratory system, and this structure was then applied by the clinical members of the group to all patients with a respiratory condition in the data set (non-surgical, n=1556; surgical, n=1374). Although the hierarchical system worked well, a three-tiered hierarchy was insufficiently detailed to code comprehensively the respiratory conditions in the data set. A four-tiered hierarchical system with the structure ‘system:site:process:condition’ was piloted on a subgroup of the respiratory system group (APACHE II diagnostic group ‘respiratory system: aspiration/poisoning/toxic’, n=149), which provided the required level of detail.

A full four-tiered coding method was then derived for all body systems, using the diagnostic information from the data set. After ensuring consistency of terminology between body systems, the method was checked to ensure that all the reasons for admission found in the five unpublished coding lists were included, and checks were made with standard textbooks to ensure that no common conditions had been missed. A fifth tier was added to code whether the reason for admission was a condition for which the patient had had surgery. This gave a final five-tier structure of ‘type: system:site:process:condition’. This was then cross-mapped to both the UK and the US APACHE II classification methods where possible (not all reasons for admission have an APACHE II code, e.g. burns).

Testing the coding method
The resulting coding method was tested initially against the data used to derive it. All 10 806 conditions in the data set were coded using the new method to ensure that there were no omissions or inconsistencies. The coding method was then tested retrospectively by coding 140 recorded reasons for admission from one general ICU, and then prospectively in three ICUs where members of the working group practised. Finally, the method was tested in five independent ICUs. At each stage the method was modified to remove inconsistencies and include any omitted reasons for admission.

The coding method was then used by all ICUs contributing data to the national audit of intensive and high-dependency care (the Case Mix Programme). Two members of staff from each unit received a half-day’s training in the background and use of the coding method, a detailed manual and ongoing support from ICNARC. To date, the method has been used for over 3 yr to code the reason for admission for 62 780 patients. User feedback has resulted in four minor revisions to date.


    Results
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
Description of the ICNARC Coding Method
The final hierarchy has five tiers: type, system, site, process and condition.

• Type: is it a surgical code (reason for surgery) or non-surgical code?

• System: which body system is involved?

• Site: which anatomical site is involved?

• Process: which physiological or pathological process is involved?

• Condition: what is the name of the condition?

Figure 1 shows how the ICM would be used to code a patient with pneumococcal pneumonia.



View larger version (30K):
[in this window]
[in a new window]
 
Fig 1 An example of how the coding method would be used to code a patient with bacterial pneumonia. The patient did not have surgery for the condition, so the first tier is non-surgical. The body system involved is the respiratory system, and the anatomical site is the lungs. The process is infection and the condition is bacterial pneumonia. The computer generates a numerical code that describes the route through the hierarchy.

 
The 12 possible body systems and 55 pathological or physiological processes are shown in Table 1. There are three systems that do not correspond to conventional anatomical or physiological body systems. The psychiatric system is used to code for mental health conditions and has no anatomical site tier. The trauma and poisoning systems are simply for user convenience when coding multiple trauma or poisoning and are duplicates of all the conditions in which the process tier is either trauma or poisoning, thus reordering the hierarchy into ‘type:process:system:site:condition’.


View this table:
[in this window]
[in a new window]
 
Table 1 The body systems and all the processes used in the ICM. The body system is qualified by the anatomical site and the process by the condition. The order of the systems in the system tier of the hierarchy was determined by the frequency of admissions in the UK APACHE II study; the most common system is at the top. Only a subset of relevant processes is presented for each anatomical site, as can be seen in Fig. 1
 
The ICM was developed to code up to five reasons for admission. These are the primary reason for admission to intensive care (corresponding to the APACHE II primary indication for ICU admission); the secondary reason for admission to intensive care; up to two other conditions relevant to the admission; and a field allowing revision of the primary reason if further information became available after the first 24 h. As a result, the ICM has to allow coding of conditions that would never be a primary reason for admission to intensive care, but might be a secondary reason. A simple example would be a patient admitted after a suicide attempt with a paracetamol overdose. The reason for admission is clearly self-poisoning with paracetamol. However, if the patient took the overdose as a result of depression and the clinician felt this was a relevant condition, this could be coded with the psychiatric system.

The ICM also contains a number of minor surgical conditions, such as inguinal hernia. These are included to allow coding of the surgical condition when patients are admitted for a complication of surgery or anaesthesia, such as persistent neuromuscular blockade as a result of pseudocholinesterase deficiency.

Specific comorbidity (chronic health history) required for the APACHE II method and the MPM IIs is recorded separately, as there are very specific requirements for both diagnostic detail and timing.

The ICM was developed to be as user-friendly as possible. As a result, conditions that affect more than one body system are coded under all the relevant systems. This occurs most commonly with neurological conditions affecting the respiratory system; for example, a patient with Guillain–Barré syndrome can be coded by two routes:

• Type: non-surgical code

• System: neurological

• Site: peripheral nervous system

• Process: inflammation

• Condition: Guillain–Barré syndrome

or

• Type: non-surgical code

• System: respiratory

• Site: peripheral nervous system disorders causing respiratory failure

• Process: inflammation

• Condition: Guillain–Barré syndrome.

This duplication of final codes means there are 1140 final codes but only 741 unique conditions. The codes corresponding to each unique condition are linked to find the overall incidence of the condition.

If a user has no clear idea of the condition that precipitated admission to intensive care, or if the condition that the patient is suffering from does not appear in the coding method, the coding can be limited to ‘type: system:site’ or ‘type:system:site:process’. If an admission cannot be coded to full depth, the user is encouraged to give the reason for admission as a text description, both to allow recognition and eventual incorporation of very rare conditions and to allow the ICM to be updated with missing, more common, conditions. The ICM was designed to be used in conjunction with a physiological data set, so non-specific syndromes, which are defined by a range of physiological abnormalities (e.g. the systemic inflammatory response syndrome), are not included as they should be determined objectively from physiological data.

Numerical results
The ICM has been used on 62 780 admissions to 118 ICUs, from which data on 22 059 admissions to 62 ICUs UK have been validated to date. Results are only given for validated data. Individual ICUs have coded between 60 and 1610 (mean 356) admissions. All but 50 (0.2%) of the admissions could be coded, 38 of the ICUs were able to code every admission and the maximum number of uncoded admissions for any unit was 7 (2.9%). Fifty (0.2%) admissions could not be coded in the first 24 h of admission but were subsequently coded when more information became available after that period. Of the admissions, 96.1% were fully coded to the full level of detail in the coding method (type:system:site:process:condition) in the primary reason for admission.

The fields required to code the admissions (from the primary reason for admission to intensive care, the secondary reason for admission to intensive care, two other conditions relevant to the admission) are shown in Table 2. Six hundred and thirty-seven of the total of 741 unique conditions (85.9%) have been used in one of the fields in the 22 059 admissions, 564 (76.1%) in the primary reason for admission. Five conditions account for 19.4% of all the primary reasons for admission. These are: aortic (abdominal) or iliac dissection or aneurysm (n=1444, 6.5%), acute myocardial infarction (n=878, 4.0%), pneumonia with no organism isolated (n=678, 3.1%), bacterial pneumonia (n=657, 3.0%) and primary brain injury (n=616, 2.8%).


View this table:
[in this window]
[in a new window]
 
Table 2 The distribution of the fields used to code the reason for admission for 22 059 admissions
 

    Discussion
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
This is first time a method for coding the reason for admission to intensive care has been developed empirically from good qualitative data. All other intensive care-specific methods have been based on expert opinion. The use of a large database of text descriptions of reasons for admission to develop and test the method ensured that all common conditions were included; this was confirmed by the very low rate (0.2%) of admissions that could not be coded and the high rate (96.1%) of fully coded admissions in the prospective evaluation. An average-sized ICU in the UK has about 500 admissions a year, so 22 059 admissions represents the equivalent of over 40 yr of admissions, a physician’s working life. Thus, any conditions not included will represent no more than a tiny fraction of any ICU’s workload.

Many of the intensive care-specific coding methods include some element of hierarchical coding. The APACHE II method divides admissions by surgical status as the first tier, then by body system as the second, and then by one of 36 precipitating factors as the third tier. The APACHE III method divides admissions by body system as the first tier and surgical status as the second, with a list of conditions as the third tier. The Fichier Diagnostique divides admissions by body system as the first tier, at the second tier by ‘symptoms, syndromes and failures’/‘illnesses and diseases’/‘chronic illnesses’, and by a mixture of anatomical sites and processes as the third and final tier. The computerized hierarchical structure developed for the ICM is far more detailed and consistent than previous methods. Unlike the APACHE II system, the ICM presents only relevant choices at each level (Fig. 1), so rapid selection of a single condition from over 700 possible options is possible. Results of independent inter-rater reliability testing5 suggest that the inter-physician agreement is good ({kappa}=0.7 for condition, 0.66 for process, 0.72 for site and 0.77 for system). Given the multiple routes to final conditions, {kappa} values will always underestimate the level of inter-observer agreement, so the agreement is probably far better than these figures suggest. Further testing is warranted.

By collecting data on reasons for admission down to the condition level using a standard, detailed, hierarchical method, the number of potential uses is increased. Any one of the 741 unique conditions, however common or rare, can be studied. In addition, data can be pooled at any level of the hierarchy, for example gastrointestinal conditions (body system), gastrointestinal tumours (body system/process) and gastric tumours (body system/site/process). Data can be pooled both from a clinical perspective (conditions that are similar clinically) or from a statistical perspective (conditions that exhibit similar characteristics, such as similar outcomes). By supplementing the reason for admission data with the sociodemographic, physiological and outcome data, important descriptive information can be obtained. These additional data allow further options for grouping, for example, grouping by sex or severity of illness for the same condition. The reason for admission can be pooled at various levels to provide sufficient sample sizes to employ statistical models to explain and predict outcome. The ICM combines a high level of detail about individual reasons for admission with a standardized, repeatable method to coalesce rare reasons for admission into groups of sufficient size for statistical analysis.

All ICUs participating in the national comparative audit of patient outcome (Case Mix Programme) have to supply at least one coded reason for admission for each patient. In addition, ICUs are asked to code any other conditions they feel are relevant to the admission, but this is not mandatory. The use of other fields is widely variable across ICUs, suggesting that opinions regarding the relevance of other conditions vary. Clearly, techniques to improve the consistency of coding for secondary and other conditions must be explored and developed.


    Acknowledgements
 
This work was undertaken at ICNARC, London, and 134 ICUs in the UK. The ICM working group comprised Dr J. Grant, Dr J. Kerr, Dr J. Morris, Dr K. Rowan, Dr A. Short, Dr S. Stone (née Michaelides) and Dr D. Young. Dr W. Brampton, Dr P. Nightingale and Dr D. Ryan assisted with prospective testing and evaluation. The authors gratefully acknowledge the 26 ICUs that participated in the Intensive Care Society’s UK APACHE II study and the 134 ICUs participating in the Case Mix Programme. This work was supported by ICNARC.


    References
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
1 Rowan KM, Kerr JH, Major E, McPherson K, Short A, Vessey MP. Intensive Care Society’s APACHE II study in Britain and Ireland—I: Variations in case mix of adult admissions to general intensive care units and impact on outcome. Br Med J 1993; 307: 972–7[ISI][Medline]

2 Rowan KM, Kerr JH, Major E, McPherson K, Short A, Vessey MP. Intensive Care Society’s APACHE II study in Britain and Ireland—II: Outcome comparisons of intensive care units after adjustment for case mix by the American APACHE II method. Br Med J 1993; 307: 977–81[ISI][Medline]

3 Rowan KM, Kerr JH, Major E, McPherson K, Short A, Vessey MP. Intensive Care Society’s Acute Physiology and Chronic Health Evaluation (APACHE II) study in Britain and Ireland: a prospective, multicenter, cohort study comparing two methods for predicting outcome for adult intensive care patients. Crit Care Med 1994; 22: 1392–401[ISI][Medline]

4 O’Neil M, Payne C, Read J. Read Codes version 3: a user led terminology. Methods Inf Med 1995; 34: 187–92[ISI][Medline]

5 de Keizer NF, Bonsel GJ, Goldfrad C, Rowan KM. The added value that increasing levels of diagnostic information provide in prognostic models to estimate hospital mortality for adult intensive care patients. Intensive Care Med 2000; 26: 577–84[ISI][Medline]