Sage Journals: Discover world-class research

Abstract

This study uses hospital administrative data to ascertain the differences in the patient characteristics, process and outcomes of care between the Emergency Department (ED) triage categories of patients admitted from an ED presentation into a large metropolitan teaching hospital with a Stroke Care Unit. Bayesian Networks (BNs) derived from the administrative data were used to provide the descriptive models. Nearly half the patients in each stroke subtype were triaged as ‘Urgent’ (to be seen within 30 minutes). With a decrease in the urgency of triage categories, the proportion admitted within 8 hours decreased dramatically and the proportion of formal discharge increased. Notably, 45% of transient ischaemic attacks (TIAs) were categorized as ‘Semi-urgent’ (to be attended within 60 minutes), indicating an opportunity to improve emergency assessment of TIAs. The results illustrate the utility of hospital administrative data and the applicability of BNs for review of the current triage practices and subsequent impact.

Keywords

administrative data Bayesian Networks emergency service stroke triage

Introduction

Stroke has been acknowledged as an important and expensive medical and societal problem.^1–5 Current stroke treatments can substantially improve the outcome; however, these therapies must be administered promptly and properly.^5–8 The crucial elements of an organized stroke care within a hospital include appropriate triaging and imaging in the emergency department (ED), prompt transfer to a dedicated stroke care unit (SCU) for appropriate and focused care, and rehabilitation.^6,7,9,10

Once signs and symptoms of stroke are recognized, effective triage for the most appropriate emergency care was listed as one of the essential features of optimal stroke care by the Brain Attack Coalition, National Institutes of Health.¹⁰ Appropriately triaged stroke patients can be fast-tracked through the ED.⁷ Since 2000, the Australian Triage Scale (ATS), based on the National Triage Scale (NTS) developed by the Australian College of Emergency Medicine (ACEM), has been in used in Australian hospitals.^11–13 The above classification reflects the urgency of medical and nursing care. It is also used as a case mix tool to monitor resource usage and efficiency.^13–15 As recommended by the ACEM, the percentage of patients seen within the time specified by each triage category is commonly used to audit and benchmark performance; with a waiting time of more that two hours considered a failure of both access and quality.¹⁴ Therefore, it would be of interest to study how the stroke patients are triaged using the ATS.

As mentioned, after appropriate treatment in ED, admission into an organized SCU provides the optimal care. The cost benefit value of SCUs has been acknowledged.^2,16,17 The benefits of SCUs in reducing in-hospital mortality, the length of stay (LOS), as well as improving the functional outcome have been reported.^18–22 In Australia, SCUs are usually found in teaching hospitals.²³ To date, there are no studies of the impact of ED triaging of suspected stroke that are also admitted to a SCU, i.e. those who have an in-patient SCU episode following an ED presentation. Such patients often use the healthcare system for a longer time and can therefore incur higher costs.

Ongoing monitoring of the care process and outcome using routinely collected administrative datasets provides opportunities to increase the efficiency of care of especially high-cost patients.^24–26 Record linking has been increasingly used to join small and large health datasets.^27–31 Linking hospital administrative datasets can present important opportunities to investigate specific questions and study the trends in the continuum of hospital care.^25,30,32 In this study, ED data linked to admitted patient data was used.

Bayesian Networks (BNs) provide a natural model for exploration and reasoning about complex probability distributions,^33–35 which is amenable to population from routinely collected administrative data. Structurally, BNs are directed acyclic graphs (DAGs), with the nodes representing random variables, the arcs representing statistical dependencies among the variables, and each node having an associated conditional probability of the node being a specific value given the values of its parents.³³ When nodes (variables) are linked in the form of a network, Bayes’ rule helps to predict the outcomes of events that are dependent on another event(s). There are many examples of health applications of BNs, including applications in nursing,^36,37 intensive care unit decision making^36–41 and public health.^42,43

Lucas³⁵ believes that BNs can be an alternative to logistic regression, with the advantage that statistical dependence and independence are explicitly represented and not hidden in approximation. It has been reported that Bayesian classifiers generally tend not to over-fit to the training data, resulting in greater tolerance to noise and thus out-performing logistic regression.^44,45 In the medical context, it has been found that Bayesian classifiers are as good, or better, in performance, ease of construction and ease of interpretation.⁴⁶ Based on statistical data and learning rules, BNs can generally improve the reliability of a model and thus be useful as an exploratory data analysis tool for studying the relationships among variables.³⁶ BNs have also been shown to provide similar performance to decision trees and to have complementary value for providing insight into care processes.⁴⁷ Modern BN tools allow one to easily navigate a wide range of options by fixing the values of specific parameters of interest and viewing the resultant conditional probability distributions.^48–50 Owing to the above factors, BN was the method of choice for the present analysis.

By applying BN to the administrative dataset of a large metropolitan teaching hospital with a SCU, comparisons were made between the admitted stroke patients who had different allocated ED triage categories. The comparisons were in terms of characteristics and care journey/outcome profiles. The study cohort was patients admitted into the same hospital as their ED presentation.

Methods

Dataset, preparation and variables

Dataset, healthcare facility and year of collection

Two fiscal years (1 June 2004–30 July 2006) of ED and in-patient data collected by a major Melbourne (Victoria, Australia) metropolitan teaching hospital with a SCU were used in this study. All hospitals in Victoria have to comply with standard definitions for reporting financial and statistical data.^25,51 Hence, the collection comprises of standardized data units required for reporting.

Record-linking and data preparation

The hospital in-patient collection data was linked with the ED data collection using deterministic matching. Deterministic matching uses unique variables in the two datasets that are to be linked and are achievable when the datasets are relatively small and fairly accurate.^29,32,52,53 Patient first name and surname, date of birth, date of presentation and discharge from ED, data of admission to ward and hospital record numbers were used in matching. After linking, a random sample of 20 records was examined and compared with the physical record for quality of link. The link was found to be correct, i.e. the process achieved the joining of records of a single patient who had an in-patient episode following an ED presentation in the same hospital.

The linked fiscal years were merged together, with great care taken to address any differences in data dictionary and International Classification of Diseases (ICD) codes. By choosing consecutive year data, in this study such heterogeneities were minimized. The linked and merged data was checked for completeness and duplicates, and any records with missing or duplicate data removed. Data dictionaries for the fiscal years (obtained from the Victorian Department Health) were used to check for errors in integrity. Exploratory data analysis was performed to check the success of the above data preparation. Lastly, the relevant data elements were transformed (see below) for use.

Cohort selection and variables of interest

All patients included in this study presented in the ED with signs and symptoms of stroke, were identified by a triage nurse and were treated as a stroke by the ED physicians. Data elements of interest include: the diagnosis codes that help define the cohort; other diagnoses; patient demographics; number of presentations at the hospital; outcome, including ED and in-patient LOS; ED referrals and separation types; and, notably, information on ED triage categories. The cohort in this study all had a principal stroke diagnoses and so were treated for stroke in both the ED and the ward.

Each patient record in the dataset had one mandatory ICD version 10 with Australian Modification (ICD-10-AM) code for the principal diagnosis and provisions to record up to 40 other ICD codes. The principal diagnosis codes were used to select the cohort and assign the stroke subtypes as follows:⁵¹ (1) stroke, not specified as haemorrhage or infarction (all I64, SnsHorI); (2) cerebral infarction (all I63, CI); (3) other nontraumatic intracranial haemorrhage (all I62, OntICH); (4) intracerebral haemorrhage (all I61, ICH); (5) subarachnoid haemorrhage (all I60, SAH); and, (6) transient cerebral ischaemic attacks and related syndromes (all G45 except G454, TIA). The current trend in clinical thinking is that there is a continuum between TIA and stroke⁵⁴ and the blurred boundaries between TIA and stroke has been discussed.⁵⁵ Therefore, in the present study, TIAs were included in the ‘stroke’ cohort; in addition, where appropriate, distinction between TIA and stroke subtypes was also made.

Comorbidity has been reported to have a significant impact on stroke outcome.^56,57 There is a shift of thinking from generic comorbidity measures to a more disease-specific complexity-of-patient-presentation construct that favours a more person-centred perspective.^58–60 The recently developed⁶¹ comprehensive stroke-specific complexity (of patient presentation) or patient complexity measure was used in BNs. Briefly, the instrument has a selection of stroke-specific ‘condition-groups’ and an associated index value based on their relative potential to contribute towards the complexity-of-stroke-patient-presentation. The convergent validity of the developed instrument had been determined using a derived ‘pseudo gold standard’. In addition, using a state administrative dataset, the alignment of the complexity score to the expected patient factors and outcomes has been established. In the current study, the index values of the stroke patient’s nonprincipal diagnoses were used to calculate the total complexity weight or the patient ‘complexity’. The SPSS Visual Bander and the BN discretization function based on entropy minimization were used to categorize all continuous variables (including age, LOS and complexity) for optimal discrimination.

The source from which patients were referred to the current ED were classified as: (1) staff from current hospital (or this hospital); (2) oneself, family and friends (or self); (3) doctors (or Dr); and, (4) all others (as ‘rest’). The number of presentations at the study hospital was classified into 1, 2–3 and 4+. ATS includes the following categories and the targeted times for attendance:^11,62 (1) Resuscitation—within 1 minute; (2) Emergency—within 10 minutes; (3) Urgent—within 30 minutes; (4) Semi-urgent—within 60 minutes; (5) Non-urgent—within 120 minutes; and, (6) Dead on arrival. LOS of patients in ED is calculated in hours and representing node in the BN is divided into: (1) up to 4 hours; (2) 4 hours and 1 minute–8 hours; (3) 8 hours and 2 minutes–12 hours; (4) 12 hours and 1 minute–18 hours; (5) 18 hours and 1 minute–24 hours; and, (6) >24 hours and 1 minute. The in-patient LOS was converted to an ordinal variable (see above). The separation types included: (1) to rehab; (2) to elderly care; (3) transfers; (4) statistical; (5) formal; (6) in-hospital death; and, (7) left against medical advice (LAMA). Note, ‘statistical separation’ is where the patient commences a new episode of care with a new care type all within a given hospital stay. Table 1 includes the description of variables/nodes used in the BNs. In the present study, the various triage categories are compared.

Table 1.

Description of variables or nodes used in the BNs

Variable	Name in BN	Variable categories	BN labels for categories
Gender	GENDER	[1] male [2] female	[1] Male [2] Female
Age groups	AGE_GROUPS	[1] 0-50 years [2] 51-65 years [3] 66-80 years [4] ≥81 years	[1] 0-50 years [2] 51-65 years [3] 66-80 years [4] >81
Number of times that the patient has presented at the health facility to date	Encounter Number	[1} 1 presentation [2] 2-3 presentations [3] > 4 presentations	[1] Upto 1 [2] 2-3 [3] ≥4
ED referral source	ED REFERRAL	[1] By current hospital [2] Self-referred [3] Referred by external doctor [4] Other sources	[1] This hospital [2] Self [3] Dr [4] Rest
Stroke subtypes at ED and stroke subtypes on admission	The two variables: STROKE SUBTYPES @ ED& STROKE SUBTYPES @ Admission	[0] Not stroke [1] Transient cerebral ischaemic attacks [2] Subarachnoid haemorrhage [3] Intracerebral haemorrhage [4] Other nontraumatic intracranial haemorrhage [5] Cerebral infarction syndromes [6] Stroke, not specified as haemorrhage or infarction	[0] Not stroke [1] TIA [2] SAH [3] ICH [4] OntICH [5] CI [6] SnsHorI.
Complexity groups	Complexity	[1] up to 3 [2] 4-9 [3] 10-15 [4] 16-21 [5] >22	[1] Up to 3 [2] 4-9 [3] 10-15 [4] 16-21 [5] ≥22
Care types	CARE TYPES	[1] Rehab [2] Elderly [3] Acute [4] Other	[1] Rehab [2] Elderly [3] Acute [4] Other
Types of separations	SEPARATION TYPES	[1] Rehab [2] Elderly [3] Transfer [4] Statistical [5] Formal [6] Death [7] LAMA	[1] Rehab [2] Elderly [3] Transfer [4] Statistical [5] Formal [6] Death [7] LAMA
Six categories of ED LOS (in hours)	ED LOS (hours)	[1] Up to 4 hours [2] 4 hours and 1 minute-8 hours [3] 8 hours and 1 minute-12 hours [4] 12 hours and 1 minute-18 hours [5] 18 hours and 1 minute-24 hours [6] >24 hours and 1 minute	[1] Up to 4 hours [2] 4 hours and 1 minute-8 hours [3] 8 hours and 1 minute-12 hours [4] 12 hours and 1 minute-18 hours [5] 18 hours and 1 minute-24 hours [6] >24 hours and 1 minute
Eightcategories of in-patient LOS (in days)	IN-PATIENT LOS (days)	[1] Up to1 day [2] 2 days [3] 3-4 days [4] 5-7 days [5] 8-10 days [6] 11-14 days [7] 15-19 days [8] ≥20 days	[1] Up to1 day [2] 2 days [3] 3-4 days [4] 5-7 days [5] 8-10 days [6] 11-14 days [7] 15-19 days [8] ≥20 days

Table 2.

The percentage of triage categories in stroke subtypes

Triage categories	Stroke subtypes
	TIA	SAH	ICH	OntICH	CI	SnsHorI
1 = Resuscitation	0.1	18.2	18.3	12.4%	4.4	3.0
2 = Emergency	6.2	27.6	24.3	15.0%	17.3	12.2
3 = Urgent	49.0	43.7	40.0	54.5	50.6	50.7
4 = Semi-urgent	44.6	9.6	16.9	17.9	27.1	33.5
5 = Non-urgent	0.2	0.9	0.5	0.2	0.6	0.8
6 = Death	0.01	0.04	0.03	0.04	0.02	0.02

Modelling with BNs

CaMML (Causal MML) BN learning tool from Monash University Data Mining Centre⁶³ was employed for the construction of BNs. CaMML is an automated causal discovery algorithm that learns the best causal structure to account for observational data using Wallace’s Minimum Message Length (MML) principle.⁶⁴ MML is said to provide a Bayesian information-theoretic metric, making a trade-off between prior probability (model complexity) and goodness of fit, thus avoiding over-fitting the data. Across a range of problems, comparative studies have shown CaMML to provide better-than or equal performance compared with other approaches.^65,66 CaMML has been successfully used⁶⁶ to learn the local structure in BNs in a particularly flexible way using either full Conditional Probability Tables (CPTs), logit models or decision trees, or any combination of these determined on a node-by-node basis (hybrid models). As the above results in a flexible learning procedure which automatically accommodates different data sizes by preferring simple local structure representations (e.g. logit models) for small data and a richer representation (decision trees and/or CPTs) for a large data, it was the method of choice. The placement of arcs between nodes, representing conditional dependency or probabilistic correlation between variables,^41,67 were algorithmically derived from training data (n = 25% of total). Netica⁶⁸ was used for BN visualization.

Results

The cohort consisted of 1892 patients who had an ED, as well as admitted episode, in the study hospital during the study period.

The BNs

Figures 2–6 give the BNs for the various triage categories. Figure 2 is for patients that were triaged as ‘Resuscitation’ (i.e. 100% Resuscitation), Figure 3 represents those were classified as ‘Emergency’ (i.e. 100% Emergency), Figure 4 represents those triaged as ‘Urgent’, Figure 5 for ‘Semi-urgent’ and Figure 6 those classified as ‘Non-urgent’.

In interpreting Netica graphs, the following should be noted (also see Table 1 and Figure 1 which give selected callouts from BNs in Figures 2 and 6):

The rectangles represent nodes (i.e. variables), with the variable name indicated at the top of the rectangle; values of the variable, with a graphical and numeric representation of the distribution of the values, are represented below the variable name.

The rectangles with a greyed area have been selected for only a specific value (see nodes to the left in Figure 1). As mentioned above, the node ‘TRIAGE’ is set to a specific value for each BN found in Figures 2–6. Also note this study is looking at data from one of the admitting hospitals (hospital X) in the cluster (i.e. ‘ADMIT Hospital’ is set to 100% hospital X—i.e. other cases have been filtered). Note the conditions of the ethics approval do not allow the identification of the hospital.

The direction of the arcs (arrows) represents influences of one variable on the distribution of another. For example, ‘GENDER’ influences ‘AGE GROUP’ which, in turn, influences ‘ENCOUNTER NUMBER’; conversely ‘GENDER’ is not modelled as being influenced by any other variable in the model and hence has no incoming arrows (see nodes with arrows on the right of Figure 1).

As can be seen from Figures 2–6, each of the nodes contains a conditional probability table that gives the probabilities of the node being in a specific state. Thus, the table contains conditional probabilities that the node is in a specific state for definite configuration of states of its parents.

Figure 1.

Selected Callouts from Bayesian Networks in Figures 2–6

Figure 2.

Bayesian Network for TRIAGE = 1 = RESUSCITATION

Figure 3.

Bayesian Network for TRIAGE = 2 = EMERGENCY

Figure 4.

Bayesian Network for TRIAGE = 3 = URGENT

Figure 5.

Bayesian Network for TRIAGE = 4 = SEMI-URGENT

Figure 6.

Bayesian Network for TRIAGE = 5 = NON-URGENT

Key differences

In the study cohort,the gender distribution between triage categories was similar, with the gender influencing age.The age differences between triage categories were also unexceptional. From the arc direction in the BNs, it can be see that the age influences the number of presentations at the hospital, the stroke subtype at admission, the type of ED referral and the separation types.

The number of current hospital encounters did not have an influence on any of the studied parameters and there was no relationship between the triage categories and encounter number.

The Percentages of Triage Categories in Stroke Subtypes are listed in Table 1. The percentage of less than or equal to eight hours ED LOS in the triage categories ‘Resuscitation’, ‘Emergency’, ‘Urgent’, ‘Semi-urgent’ and ‘Non-urgent’ were 72.8%, 60.2%, 51.3%, 43.4% and 40.6% respectively.

In-patient diagnosis influenced the provisional diagnosis at ED presentation.There was no significant difference in the complexity of patient presentation between triage categories. However, stroke subtype (at admission) influenced the complexity of presentation and, in turn, patient complexity affected in-patient LOS. The proportion of patients who admitted to the same hospital within eight hours decreased dramatically as the ED triage category changed from 1 to 5.

In the study cohort, 19.9% of those who received resuscitation were transferred to other facilities and 62.1% who underwent resuscitations in ED died after admission. The in-hospital death declined with the lower priority triage categories. Eighty percent of the patients classified in ED as ‘Emergency’ were either transferred to other facilities (43.0%) or formally discharged (37.0%) from the hospital. The proportion of formal discharge from hospital was 54.6% for ‘Urgent’, 64.8% for ‘Semi-urgent and 60.8% for ‘Non-urgent. For the triage categories 3–5, the proportion of transfers was between 32.3% and 36.8%.

Discussion

Stroke and transient TIA are increasingly common conditions that are associated with significant morbidity and mortality.⁶⁹ The triage scale is said to relate to a number of outcome measures and provides an opportunity for operational assessment.¹⁴ BNs were used here to analyze hospital administrative data in order to understand the impact of ED triaging of suspected stroke patients who subsequently experienced an in-patient admission in a large metropolitan teaching hospital with a SCU. A number of parameters were compared.

The discrepancies in the ED treatment of men and women have been documented, with a pattern of delayed care and poorer outcomes for women.⁷⁰ However, in this study the gender differences between triage categories were unremarkable, with gender having an impact on age distribution. The latter could be explained by the larger proportion of females in the older age groups.

Although there were no age differences between triage categories, age was found to influence the number of presentations at the hospital, the stroke subtype at admission, the type of ED referral and the separation types. The influence of age is not surprising as the elderly frequent the hospital more often and stroke is more frequent in the older age groups.³

There was no relationship between the triage categories of suspected stroke and the number of visits to the current hospital. Surprisingly, the number of presentations at this hospital does not have an influence on any of the studied parameters.

As expected, the confirmed in-patient diagnosis influences the provisional diagnosis at ED presentation. It was noted that, in ED, some non-stroke patients are incorrectly (and presumably symptomatically) diagnosed as strokes and some strokes classified as non-strokes. This was corrected at admission following full investigation. The degree of difficulty of triage and diagnosis depends on whether symptoms are specific to a particular disease.

Early recognition and intervention in stroke are required to minimize mortality and morbidity.⁷ In the current study, nearly half of the patients in each of the stroke subtypes were earmarked to be seen within 30 minutes, i.e. triaged as ‘Urgent’ (see Table 1). A study of EDs in Houston (TX, USA)⁷¹ reported that the average time taken between ED arrival to a physician examination and computed tomography (CT) scan was 28 and 100 minutes respectively. From their review of the literature, Kwan et al.⁷² found that the median delay from arrival at hospital to medical assessment ranged from 20 minutes to 4 hours. The above systematic review warns that, unless strokes are triaged as high priority to be assessed and managed quickly (ideally within one hour), thrombolysis is not likely to have a major effect on outcome. With the study cohort, it would be of interest to investigate the time to examination, CT scan and treatment. Although SCUs and thrombolysis have greatly improved the management of stroke patients, the benefits of these advances can only be fully realized if emergency departments recognize the signs of stroke and act both appropriately and promptly.^6,7 After reviewing the literature, an Australian study⁷³ concluded that factual knowledge was more important than the years of emergency nursing or triage experience when it comes to triage decisions.

In the study cohort, 45% of TIA, 35.5% of SnsHorI, 27% of CI, nearly 18% of OntICH, nearly 17% of ICH and less that 10%of SAH were categorized as ‘Semi-urgent’ (to be attended within 60 minutes). TIAs are known to be an important risk factor for stroke.^74,75 As nearly half of all post-TIA recurrent strokes is said to occur in the first 24 hours, there is a need for emergency assessment of TIAs.⁷⁶

The assessment at ED is done quickly (usually 2–5 minutes), with the patients’ triage being based on the need for time-critical intervention rather than severity of the illness.¹⁴ Triage nurses are said to make valid and reliable estimates of patient complexity.⁷⁷ In the present study, the differences in the complexity of patient presentation between triage categories were not noteworthy—perhaps because triage is for a suspected stroke rather than patient complexity. Stroke subtype (at admission) influences the complexity of presentation. Patient complexity (and not the stroke subtype) impacts in-patient LOS.

The efficiency of an ED depends on many factors, including the delay in handing over to the appropriate hospital wards/units/departments for continuing care. Such access block and its effects on patient care are known.^78–80 Therefore, to understand (and improve) ED LOS, one has to also focus on the area of the hospital that provides the next episode of care. An Australian study reported that the patients kept in the ED for longer than eight hours had an average in-patient LOS of 4.9 days compared with 4.1 days for the other patients.⁸¹ The observed correlation between prolonged ED stay and excess in-patient LOS remained, even when results were adjusted for major confounding factors.⁸⁰ In the present study, the percentage of stroke patients presenting at ED who were admitted in the same hospital within eight hours were examined. The proportion admitted within eight hours decreased dramatically as the ED triage category changed from 1 to 5. ED delays and, hence, delay to treatment can be detrimental for patients with especially time-critical conditions.⁸² A relationship between time to treatment and clinical outcomes has been demonstrated for patients with acute ischaemic stroke⁸³ and in acute stroke and TIA.⁸⁴ As timely and accurate imaging done in an ED guides the next course of action and outcome,⁸⁵ it would be informative to study the time to imaging in ED and on-going treatment.

Stroke is a medical emergency that needs prompt attention to avoid the possibility of severe deficits, complications and death.^10,84,86,87 In the study cohort, 62.1% who receive resuscitation in ED died as an in-patient and 19.9% were transferred to other facilities. The percentage of in-hospital death decreased with the lower priority triage categories. The majority of patients classified at ED as ‘Emergency’ are either transferred to other facilities (43.0%) or formally discharged (37.0%) from the hospital. The proportion of formal discharge from hospital was observed to increase with a decrease in the urgency of triage categories (54.6% for ‘Urgent’, 64.8% for ‘Semi-urgent and 60.8% for ‘non-urgent’). The proportions of transfers remained at similar level (32.3–36.8%) for the triage categories 3–5. It appears the triaging practice at this hospital appears to be generally appropriate and achieves the desired effect.

Conclusions

BNs were used in this study as this method, via a graphical diagram, provides a comprehensive way of representing relationships and influences among nodes or variables. Thus, BNs are a convenient tool to model and subsequent display and explore associations among a set of variables.This study demonstrated that the selected algorithm and the hybrid BNs, when applied to stroke patient’s who were admitted from an ED presentation, can give an overview of the differences in the stroke patient journey/outcome profiles between the ED triage categories.

In the cohort of stroke patients who had presented at the ED of a major metropolitan teaching hospital with SCU and continued as an admitted episode, overall many of the characteristics and outcomes are as expected, indicating a good performance. Nearly half of the patients in each of the stroke subtypes was triaged as ‘Urgent’, to be seen within 30 minutes. The proportion admitted within eight hours decreased dramatically as the ED triage category changed from 1 to 5; the proportion of formal discharge from hospital was observed to increase with a decrease in the urgency of triage categories.It is notable, however, that 45% of TIAs were categorized as only ‘Semi-urgent,’ indicating an opportunity to improve emergency assessment of TIAs.In this study, unlike other studies, the gender differences between triage categories were unremarkable.

There are some limitations in this study. As with any analysis/modelling, the results can be influenced by the quality of the data collection and the computational methodology employed in the construct of BNs. The advantage of using the standardized hospital administrative data has already been pointed out and BNs are a known, rigorous, well-defined technique. The BN constructs from the hospital administrative data helped to identify only a limited set of factors and gave a high-level view of process and outcome of care. Further follow-up of pre- and post-study journeys and abstraction of greater details from the medical records would be required to better explain some of the observed differences.

Emergency departments on their own, as well as the ED to in-patient journey, present unique challenges for process improvement. Studies such as this are important for gathering information on the impact of current triage practices and highlighting areas for further exploration.

Footnotes

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

References

Cadilhac

. The changing face of stroke. Medical management revolution. Aust Nurs J 1998; 5: 18–21.

Saka

Serra

Samyshkin

McGuire

Wolfe

. Cost-effectiveness of stroke unit care followed by early supported discharge. Stroke 2009; 40: 24–29.

Donnan

Fisher

Macleod

Davis

. Stroke. Lancet 2008; 371: 1612–1623.

Crimmins

Levi

Gerraty

Beer

Hill

. Acute stroke and transient ischaemic attack management—time to act fast. Intern Med J 2009; 39: 325–331.

Young

Tolentino

. Stroke evaluation and treatment. Top 2009; 16: 389–410.

Quain

Parsons

Loudfoot

Spratt

Evans

Russell

. Improving access to acute stroke therapies: A controlled trial of organised pre-hospital and emergency care. Med J Aust 2008; 189: 429–433.

Lawson

Gibbons

. Acute stroke management in emergency departments. Emerg Nurse 2009; 17: 30–34.

Rudd

Hoffman

Irwin

Lowe

Pearson

. Stroke unit care and outcome: Results from the 2001 National Sentinel Audit of Stroke (England, Wales, and Northern Ireland). Stroke 2005; 36: 103–106.

Anonymous. A systems approach to immediate evaluation and management of hyperacute stroke. Experience at eight centers and implications for community practice and patient care. The National Institute of Neurological Disorders and Stroke (NINDS) rt-PA Stroke Study Group. Stroke 1997; 28: 1530–1540.

10.

Dion

. Management of ischemic stroke in the next decade: stroke centers of excellence. J Vasc Interv Radiol 2004; 15: S133–S141.

11.

Australasian College for Emergency Medicine (ACEM): http://www.acem.org.au/open/documents/triageguide.htm (accessed Month Year).

12.

MIHSR. Consistency of triage in Victoria’s emergency departments: guidelines for triage education and practice. Clayton: Monash Institute of Health Services Research (MIHSR), 2001.

13.

Yousif

Bebbington

Foley

. Impact on patients triage distribution utilizing the Australasian Triage Scale compared with its predecessor the National Triage Scale. Emerg Med Australas 2005; 17: 429–433.

14.

Australian Collage of Emergency Medicine (ACEM): http://www.acem.org.au/infocentre.aspx?docId=59#GUIDELINES (accessed Month Year).

15.

Fry

Rhodes-Sutton

. A retrospective chart review of adult mortality characteristics of patients presenting to a principal tertiary emergency department. Accid Emerg Nurs 2005; 13: 122–125.

16.

Fuentes

Diez-Tejedor

. Stroke units: Many questions, some answers. Int J 2009; 4: 28–37.

17.

Launois

Giroud

Megnigbeto

Le Lay

Presente

Mahagne

. Estimating the cost-effectiveness of stroke units in France compared with conventional care. Stroke 2004; 35: 770–775.

18.

Stroke Unit Trialists Collaboration. Organised inpatient (stroke unit) care for stroke. Cochrane DatabaseSyste Rev 2007: CD000197.

19.

Diez-Tejedor

Fuentes

. Acute care in stroke: Do stroke units make the difference? Cerebrovasc Dis 2001; 11(Suppl. 1): 31–39.

20.

Zhu

Newcommon

Cooper

Green

Seal

Klein

. Impact of a stroke unit on length of hospital stay and in-hospital case fatality. [Calgary Stroke Program]. Stroke 2009; 40: 18–23.

21.

Indredavik

Fjaertoft

Ekeberg

Loge

Morch

. Benefit of an extended stroke unit service with early supported discharge: A randomized, controlled trial. Stroke 2000; 31: 2989–2994.

22.

Counsell

Warlow

Sandercock

Fraser

van Gijn

. The Cochrane Collaboration Stroke Review Group. Meeting the need for systematic reviews in stroke care. Stroke 1995; 26: 498–502.

23.

Read

Levy

. Differences in stroke care practices between regional and metropolitan hospitals. Intern Med J 2005; 35: 447–450.

24.

Luthi

Troillet

Eisenring

Sax

Burnand

Quan

. Administrative data outperformed single-day chart review for comorbidity measure. Int J Qual Health Care 2007; 19: 225–231.

25.

Nadathur

. Maximising the value of hospital administrative datasets. Aust Health Rev 2010; 34: 1–8.

26.

Malenka

McLerran

Roos

Fisher

Wennberg

. Using administrative data to describe casemix: A comparison with the medical record. J Clin Epidemiol 1994; 47: 1027–1032.

27.

Williams

Dobb

Finn

Knuiman

Lee

Geelhoed

. Data linkage enables evaluation of long-term survival after intensive care. Anaesth Intensive Care 2006; 34: 307–115.

28.

Riley

. Administrative and claims records as sources of health care cost data. Med Care 2009; 47(Suppl. 1): S51–S55.

29.

Boyle

. The experience of linking Victorian emergency medical service trauma data. BMC Med Inf Decis Mak 2008; 8: 52.

30.

Cameron

Purdie

Kliewer

McClure

. Differences in prevalence of pre-existing morbidity between injured and non-injured populations. Bull World Health Organ 2005; 83: 345–352.

31.

Potosky

Riley

Lubitz

Mentnech

Kessler

. Potential for cancer related health services research using a linked Medicare-tumor registry database. Med Care 1993; 31: 732–748.

32.

Victor

Mera

. Record linkage of healthcare insurance claims. Stud Health Technol Inform 2001; 84: 1409–1013.

33.

Nadathur

. Bayesian Networks in the Health Domain. In: Ali

Xiang

. (eds) Dynamic and Advanced Data Mining for Progressing Technological Development. Hershey: IGI Global Publication, 2009, pp. 342–376.

34.

Lucas

Van der Gaag

Abu-Hanna

. Bayesian networks in biomedicine and health. Artif Intell Med 2004; 30: 201–214.

35.

Lucas

. Bayesian analysis, pattern analysis, and data mining in health care. Curr Opin Crit Care 2004; 10: 399–403.

36.

Lee

Abbott

. Bayesian Networks for knowledge discovery in large datasets: Basics for nurse researchers. J Biomed Inf 2003; 36: 389–399.

37.

Tervo-Heikkinen

Kiviniemi

Partanen

Vehvilainen-Julkunen

. Nurse staffing levels and nursing outcomes: a Bayesian analysis of Finnish-registered nurse survey data. J Nurs Manag 2009; 17: 986–993.

38.

Smith

Doctor

Meyer

Kalet

Phillips

. A decision aid for intensity-modulated radiation-therapy plan selection in prostate cancer based on a prognostic Bayesian network and a Markov model. Artif Intell Med 2009; 46: 119–130.

39.

Sierra

Serrano

Larranaga

Plasencia

Inza

Jimenez

. Using Bayesian networks in the construction of a bi-level multi-classifier. A case study using intensive care unit patients data. Artif Intell Med 2001; 22: 233–248.

40.

Abston

Pryor

Haug

Anderson

. Inducing practice guidelines from a hospital database. Proceedings of the AMIA Annual Fall Symposium 1997: 168–172.

41.

Bidyuk

Terent’ev

Gasanov

. Construction and methods of learning of Bayesian Networks. Cybernetics and Systems Analysis 2005; 41: 587–598.

42.

Donald

Cook

Mengersen

. Bayesian network for risk of diarrhea associated with the use of recycled water. Risk Anal 2009; 29: 1672–1685.

43.

Mnatsakanyan

Burkom

Coberly

Lombardo

. Bayesian information fusion networks for biosurveillance applications. J Am Med Inform Assoc 2009; 16: 855–863.

44.

Cloutier

Sirois

. Bayesian versus Frequentist statistical modeling: A debate for hit selection from HTS campaigns. Drug Discov Today 2008; 13: 536–542.

45.

Klon

Lowrie

Diller

. Improved naive Bayesian modeling of numerical data for absorption, distribution, metabolism and excretion (ADME) property prediction. J Chem Inf Model 2006; 46: 1945–1956.

46.

Van der Gaag

Renooij

Feelders

de Groote

Eijkemans

MJC

Broekmans

. Aligning Bayesian Network classifiers with medical contexts. In: Technical Report UU-CS-2008-015. Utrecht: Department of Information and Computing Sciences, Utrecht University, 2008.

47.

Bayat

Cuggia

Rossille

Kessler

Frimat

. Comparison of Bayesian network and decision tree methods for predicting access to the renal transplant waiting list. Stud Health Technol Inform 2009; 150: 600–604.

48.

Van Gerven

MAJ

Taal

Lucas

. Dynamic Bayesian networks as prognostic models for clinical patient management. J Biomed Inf 2008; 41: 515–529.

49.

Curtis

Cheng

Delzell

Fram

Kilgore

Saag

. Adaptation of Bayesian data mining algorithms to longitudinal claims data: Coxib safety as an example. Med Care 2008; 46: 969–975.

50.

Verduijn

Peek

Rosseel

PMJ

De J

De Mol

BAJM

. Prognostic Bayesian Networks. J Biomed Inf 2007; 40: 609–618.

51.

Nadathur

. Recorded categories of non-principal diagnoses in Victorian public hospital transient ischaemic attack and stroke admissions. Health Information Management Journal 2006; 37: 33–44.

52.

Roos

Wajda

. Record linkage strategies. Part I: Estimating information and evaluating approaches. Methods Inf Med 1991; 30: 117–123.

53.

Gomatam

Carter

Ariet

Mitchell

. An empirical comparison of record linkage procedures. Stat Med 2002; 21: 1485–1496.

54.

Sacco

Adams

Albers

Alberts

Benavente

Furie

. Guidelines for prevention of stroke in patients with ischemic stroke or transient ischemic attack: a statement for healthcare professionals from the American Heart Association/American Stroke Association Council on Stroke: co-sponsored by the Council on Cardiovascular Radiology and Intervention: the American Academy of Neurology affirms the value of this guideline. Circulation 2006; 113: e409–e449.

55.

Mitka

. Rapid diagnosis and treatment of TIAs help reduce recurrent stroke risk. AMA 1992;267: 2197–203.

56.

Kazmierski

. Predictors of early mortality in patients with ischemic stroke. Expert Review of Neurotherapeutics 2006; 6: 1349–1362.

57.

Fischer

Arnold

Nedeltchev

Schoenenberger

Kappeler

Hollinger

. Impact of comorbidity on ischemic stroke outcome. Acta Neurol Scand 2006; 113: 108–113.

58.

Valderas

Starfield

Sibbald

Salisbury

Roland

. Defining comorbidity: Implications for understanding health and health services. Ann Fam Med 2009; 7: 357–363.

59.

Elixhauser

Steiner

Harris

Coffey

. Comorbidity measures for use with administrative data. Med Care 1998; 36: 8–27.

60.

Bonavita

De Simone

. Towards a definition of comorbidity in the light of clinical complexity. Neurol Sci 2008 24: 65–9.

61.

Nadathur

. An ICD-10 based complexity-of-patient-presentation measure for stroke. 2011; (submitted).

62.

Victorian Department of Health (VDoH), Melbourne, Australia: http://www.health.vic.gov.au/hdss/vemd/index.htm (accessed Month Year).

63.

http://www.datamining.monash.edu.au/ (accessed Month Year).

64.

Wallace

Korb

. Learning linear causal models by MML samplling. In: Gammerman

. (ed.) Causal models and intelligent data management. Berlin, New York: Springer, 1999.

65.

Dai

Korb

Wallace

. A study of causal discovery with weak links and small samples. In: Proceedings of the 15th International Joint Conference on Artificial Intelligence. Morgan Kaufmann 1997,pp. 1304–1309.

66.

O Donnell

Allison

Korb

. Learning Hybrid Bayesian Networks by MML. Iswc International Workshops on Uncertainty Reasoning for the Semantic Web, Ursw. 2006: 192–203.

67.

Kim

I-C

Jung

Y-G

. Using Bayesian Networks to analyze medical data. In: Perner

Rosenfeld

. (eds) Proceedings of the third international conference of Machine Learning and Data Mining (MLDM). Leipzig: Springer-Verlag, 2003, pp. 317–327.

68.

http://www.norsys.com (accessed Month Year).

69.

Ovbiagele

. The emergency department: first line of defense in preventing secondary stroke. Acad Emerg Med 2006; 13: 215–222.

70.

Washington

Bird

. Sex differences in disease presentation in the emergency department. Ann Emerg Med 2002; 40: 461–463.

71.

Bratina

Greenberg

Pasteur

Grotta

Bratina

Greenberg

. Current emergency department management of stroke in Houston, Texas. Stroke 1995; 26: 409–414.

72.

Kwan

Hand

Sandercock

. A systematic review of barriers to delivery of thrombolysis for acute stroke. Age Ageing 2004; 33: 116–121.

73.

Considine

Botti

Thomas

. Do knowledge and experience have specific roles in triage decision-making? Acad Emerg Med 2007; 14: 722–726.

74.

Sacco

. Risk factors for TIA and TIA as a risk factor for stroke. Neurology 2004; 62(Suppl. 6): S7–S11.

75.

Johnston

Nguyen-Huynh

Schwarz

Fuller

Williams

Josephson

. National Stroke Association guidelines for the management of transient ischemic attacks. Ann Neuro. 2006; 60: 301–313.

76.

Chandratheva

Mehta

Geraghty

Marquardt

Rothwell

. Population-based study of risk and predictors of stroke in the first few hours after a TIA. Neurology 2009; 72: 1941–1947.

77.

Vance

Sprivulis

. Triage nurses validly and reliably estimate emergency department patient complexity. Emerg Med Australas 2005; 17: 382–386.

78.

Cameron

Campbell

. Responses to access block in Australia: Royal Melbourne Hospital. Med J Aust 2003; 178: 109–110.

79.

Cameron

Campbell

. Access block: problems and progress. Med J Aust 2003; 178: 99–100.

80.

Liew

Kennedy

. Emergency department length of stay independently predicts excess inpatient length of stay. The Medical Journal of Australia 2003; 179: 524–526.

81.

Richardson

. The access-block effect: relationship between delay to reaching an inpatient bed and inpatient length of stay. The Medical Journal of Australia 2002; 177: 492–495.

82.

Duke

Green

Briedis

. Survival of critically ill medical patients is time-critical. Critical care and resuscitation : Journal of the Australasian Academy of Critical Care Medicine 2004; 6: 261–267.

83.

Tilley

Lyden

Brott

Levine

Welch

. Total quality improvement method for reduction of delays between emergency department admission and treatment of acute ischemic stroke. The National Institute of Neurological Disorders and Stroke rt-PA Stroke Study Group. Arch Neurol 1997; 54: 1466–1474.

84.

Wester

Radberg

Lundgren

Peltonen

. Factors associated with delayed admission to hospital and in-hospital delays in acute stroke and TIA: a prospective, multicenter study. Seek- Medical-Attention-in-Time Study Group. Stroke 1999; 30: 40–48.

85.

Mullins

. Modern emergent stroke imaging: Pearls, protocols, and pitfalls. Radiol Clin North Am 2006; 44: 41–62.

86.

Rapp

Bratina

Barch

Braimah

Daley

Donnarumma

. Code Stroke: Rapid transport, triage and treatment using rt-PA therapy. The NINDS rt-PA Stroke Study Group. J Neurosci Nurs 1997; 29: 361–366.

87.

Brott

Bogousslavsky

. Treatment of acute ischemic stroke. N Engl J Med 2000; 343: 710–722.

Emergency department triaging of admitted stroke patients—A Bayesian Network analysis

Abstract

Keywords

Introduction

Methods

Dataset, preparation and variables

Dataset, healthcare facility and year of collection

Record-linking and data preparation

Cohort selection and variables of interest

Modelling with BNs

Results

The BNs

Key differences

Discussion

Conclusions

Footnotes

References