Triage of the child with abdominal pain: A clinical algorithm for emergency patient management

Abstract

OBJECTIVE:

To create a simplified clinical algorithm for the triage of children with abdominal pain.

DESIGN:

Retrospective analysis.

SETTING:

Emergency room at the Children’s Hospital of Eastern Ontario, Ottawa, Ontario.

METHODS:

A data mining methodology (rough sets analysis) was applied to a randomized data set obtained from 175 emergency room admission charts of patients. Patients were placed into two diagnostic decision classes: appendicitis confirmed by a pathological report, and resolution (this classification implied the resolution of all clinical complaints and physical findings, with no pathological diagnosis and no operative procedure).

RESULTS:

Nine clinical symptoms and signs were identified as being important in the management of children with abdominal pain. A clinically based algorithm for the triage of such children was developed.

CONCLUSIONS:

It is possible to develop a clinical algorithm for the triage of children with abdominal pain that can also be used by nonmedical professionals. A template for such an algorithm can be used as the basis for diagnosing other paediatric emergencies, such as chest pain, headaches and joint pain.

Despite the advent of ultrasonography, computer assisted tomography and magnetic resonance technology, clinical assessment remains the major diagnostic tool in paediatric emergencies. In the management of paediatric abdominal pain in the emergency room (ER), where most children with abdominal pain are referred, rapid diagnosis depends on the clinical acumen and experience of the caregiver. The presence of other patients with more acute problems and a large number of patients in the ER delay individual patient assessment. In a teaching hospital, the initial evaluation is performed by a physician in training. If the physician is not able to make a diagnosis, investigations that may further defer definitive management may be ordered. Complex radiological investigations, such as abdominal ultrasound, are ‘operator dependent’ (ie, they require the presence of an experienced radiologist). The frustration of parents and the anxiety of the patient may complicate management further.

Fiscal responsibilities are forcing medical institutions and physicians to curtail the costs of diagnostic processes. The prevailing view is that the costs of medical care can be partially controlled through innovative approaches to health care management (1). Research involving health care management shows that the usefulness of information provided by tests and procedures diminishes as diagnostic certainty is approached (2). This observation has prompted the development of measures for the evaluation, and rationalization of tests and diagnostic procedures (3). The measures either deal with proposals for organizational change in institutions that provide care (4) or they focus on new, usually information-based approaches to patient management (5). The study described in this paper falls into the second category.

Evidence from both clinical and psychological studies (6) points to the obvious advantage of the rapid triage (ie, initial assessment) of patients with abdominal pain. Children with abdominal pain constitute a significant proportion of paediatric emergencies. The central difficulty related to the triage of these children is determining the clinical symptoms and signs (attributes) that, in combination, contribute the most to the diagnosis and management of abdominal pain. A reduced set of pertinent attributes would assist the triage nurse and the ER physician.

The present paper discusses the identification of such a set of attributes and, based on the attributes, the development of a clinical algorithm to facilitate the triage of the child with abdominal pain. Improving triage decision making may also achieve the secondary goals of parental satisfaction with medical care, improved patient compliance and a reduction in the overall costs of medical services.

PATIENTS AND METHODS

A retrospective analysis of the ER records of 175 children with abdominal pain admitted to the Children’s Hospital of Eastern Ontario, Ottawa, Ontario in 1997 was completed. Based on the final diagnostic decision, patients were classified into two distinct groups (ie, decision classes):

• patients with appendicitis confirmed by pathology; and

• resolution (this classification implied the resolution of all clinical complaints and physical findings, with no pathological diagnosis and no operative procedure).

To maintain a homogenous data set, patients were excluded from the analysis if the final diagnostic decision class was neither appendicitis nor resolution. Table 1 shows the distribution of data among the two groups.

Table 1:

Final diagnostic decision class of 175 children with acute abdominal pain

Diagnostic decision class	Number of patients (%)
Resolution	139 (79.4)
Appendicitis	36 (20.6)
Total	175 (100)

For each patient, an attempt was made to collect information about 12 clinical symptoms and signs, but this was not always possible. Table 2 shows the information that was collected and the information that was missing from patients’ charts.

Table 2:

Incidence of failure to record the clinical attributes in each of two diagnostic decision classes in 175 paediatric emergency room patient records

Attribute	Diagnostic decision class
	Resolution	Appendicitis
Age of child	0	0
Sex	0	0
Abdominal pain duration	0	0
Abdominal pain site	1	0
Abdominal pain type	31	14
Vomiting (number of times occurred)	0	0
Bilious vomit	4	0
Previous visits to emergency room in past 48 h	12	0
Temperature	0	0
Abdominal site tenderness	1	1
Abdominal mass	2	2
White blood cell count	3	1

The data set used contained both qualitative and quantitative information. Comprehensive analysis of such data can be accomplished with the help of new methodologies that were developed for knowledge discovery applications. The discipline of knowledge discovery encompasses several fields, such as artificial intelligence, intelligent information systems, operations research and knowledge acquisition in expert systems. Numerous powerful knowledge discovery theories that extract patterns from imprecise data have been developed. Rough sets (RS) analysis is one such theory (7). RS theory, its mathematical foundations, basic concepts and applications have been described in the literature, including research presented by Pawlak (7), Grzymala-Busse (8), and Lin and Cercone (9). RS analysis has been used successfully to analyze Canadian medical data (10) and several other data mining problems (9). RS analysis was used to analyze the data described in this paper.

RS theory is based on the observation that it is very difficult to describe properly the characteristics of a problem when relying on imprecise information about the values of the problem’s attributes (eg, clinical symptoms and signs). In other words, imprecise information in ER records causes indiscernibility in the classification of patients into appendicitis or resolution decision classes when the final diagnostic decision is excluded. RS theory provides a powerful tool to identify a minimal subset of attributes (a reduct) that gives a satisfactory description of a decision-making problem.

The present study describes an application of RS theory to identify a reduct that enabled the classification of a patient (with an unknown final diagnostic decision) as having either appendicitis or resolution. Before data were analyzed, each patient’s chart was reviewed. Selected clinical attributes, the pathological report and/or the final clinical diagnosis were recorded. This information was tabulated and it is presented in Table 3. The rows of Table 3 represent individual charts, while the columns correspond to the attributes that were extracted from patients’ charts.

Table 3:

Example of a table used to present information about abdominal pain attributes extracted from patients’ emergency room records and to classify patients into one of two diagnostic decision classes: resolution or appendicitis

Age (years)	Sex	AbdPainDur	AbdPainSite	AbdPainType	Vomiting	Vombile	PrevVis	Temp (°C)	AbdTend	AbdMass	WBC (×103/mm3)	Class
<2	F	<4 h	Else	Interm	1	No	No	37–39	Abs	No	10–20	Resolution
2–7	M	4–24 h	RLQ	Cont	<3	Yes	No	<39	Pres	Yes	<20	Appendicitis

Rows represent individual charts and columns correspond to attributes. AbdMass Abdominal mass; AbdPainDur Duration of abdominal pain; AbdPainSite Location of abdominal pain; AbdPainType Type of abdominal pain; AbdTend Site tenderness; Abs Absent; Cont Continuous; Else Abdominal pain excluding right lower quadrant; F Female; Interm Intermittent; M Male; Pres Present; PrevVis Previous visit to the emergency room in the past 48 h; RLQ Right lower quadrant; Temp Temperature; Vombile Bilious vomit; Vomiting The number of times vomiting occurred; WBC White blood cell count

Based on RS terminology, a complete table used during this study involved a set of diagnostic decisions that consisted of ‘appendicitis’ and ‘resolution’, and a set of clinical attributes that consisted of age, sex, duration of abdominal pain, location of abdominal pain, type of abdominal pain, the number of times vomiting occurred, the presence of bile in the vomitus, whether a previous visit was made to the ER in the past 48 h, fever, site tenderness, abdominal mass and white blood cell count (WBC). Value domains for the attributes were expressed as, for example, age equals younger than two years, from two to seven years of age and from seven to 16 years of age (Age={<2y; 2–7y; 7–16y}). The values of all the attributes were discretized and coded according to their significance to clinical judgment.

A common problem associated with the analysis of medical data is how to treat missing values in patients’ charts. The authors dealt with this problem by considering each piece of missing information as a separate value and, thus, assigning it a unique artificial value (see Table 2 for information about missing values).

The application of RS allowed the authors to identify a reduct that was used to develop the clinical algorithm. The purpose of the algorithm was to describe, in an easily understood format, the causal dependencies that lead toward the classification of a patient as having either appendicitis or resolution. In the development of the clinical algorithm, the authors relied on information provided by interesting rules that were generated using RS methodology. (Interesting rules possess certain properties, such as the rules’ strength [the minimal number of historical cases that are correctly classified]. The authors assumed that a rule was considered to be an interesting rule only if it corrently classified at least 14% of patients when applied to historical data from the patients’ ER records). All calculations were conducted on an IBM-compatible computer using ProFit software (ProSoft, Poland) (11).

RESULTS

The reduction of the original set of 12 clinical symptoms and signs resulted in the generation of four reducts. Three of the reducts consisted of nine attributes, and one of the reducts consisted of 10 attributes. The longer reduct was discarded; the three shorter reducts are presented in Table 4. The last row of Table 4 shows the result of classification tests that classified patients into one of the decision classes with the help of the attributes that comprised a reduct. The classification accuracy is given as the average value obtained after a series of 10-fold-cross-validation tests. (The initial data set was divided into 10 disjoint sets that are called folds. The learning phase [ie, a phase where RS analyze the data] and the classification phase [ie, a phase where RS are used to classify ‘new’ data] were conducted 10 times, with each of the folds acting as the testing sample, and all of the remaining folds acting as the learning sample. This type of test is normally used with medium sized data sheets.) The decision rules used for the classification were generated using a minimal covering algorithm (8). (This algo-righm was designed to generate a minimal number of rules that would classify all objects [patients] on the basis of the analyzed data set [the ER charts]). The highest accuracy was achieved for reduct number 3, and it was selected as the basis for further investigation.

Table 4:

Clinical attributes that comprised reducts 1 through 3

Clinical symptoms and signs	Reduct number 1	Reduct number 2	Reduct number 3
Age	√	√	√
Sex	√	√	√
Abdominal pain duration	√	√	√
Abdominal pain site			√
Abdominal pain type	√	√	√
Vomiting (number of times occurred)	√	√	√
Presence of bilious vomit	√	√	√
Previous visits to emergency room in past 48 h	√	√	√
Temperature	√	√	√
Abdominal site tenderness		√
Abdominal mass	√
White blood cell count	√	√	√
Accuracy (mean±SD)	73.5±2.0	73.5±1.9	77.3±1.0*

^*The highest classification accuracy (into the appendicitis or the resolution diagnostic decision class) was achieved for reduct number 3, and it was selected as the basis for designing a clinical algorithm

The causal relationship between the patient’s final diagnostic decision class and reduct number 3 was expressed in terms of the interesting rules from which the clinical algorithm was derived. Reduct number 3 is comprised of the following clinical attributes: age, sex, duration of abdominal pain, location of abdominal pain, type of abdominal pain, the number of times that vomiting occurred, previous visits to the ER in past 48 h, fever and WBC. The purpose of this algorithm is to rapidly and reliably triage a patient as having either appendicitis or a resolution. Table 5 presents the average values of the classification accuracy of the interesting rules obtained after 50 passes of 10-fold-cross-validation tests.

Table 5:

Average values of the classification accuracy (into appendicitis or resolution of symptoms diagnostic decision classes) of the interesting rules obtained after 50 passes of 10-fold-cross-validation tests

Decision class	Correct*	Incorrect*	Unable to classify*
Total	65.9±2.3	33.3±2.3	0.8±0.9
Resolution	64.6±3.2	33.6±3.2	0.8±1.0
Appendicitis	67.0±4.7	32.2±4.5	0.7±1.4

^*Data are given as mean ± SD

The total classification accuracy of the interesting rules is slightly lower than that given for a minimal covering algorithm in Table 4 because the interesting rules are more general and, thus, they represent broader classification patterns that may not be applicable to very specific cases. No significant difference was found in the classification accuracy of the interesting rules within the resolution and the appendicitis classes.

The interesting rules are not diagnostic (ie, they do not indicate the patient’s final diagnosis), but they capture general patterns of medical information associated with the assessment of the child with abdominal pain. These patterns are written in the form of a clinical algorithm. Following a format accepted by the medical profession, such an algorithm can be expressed as a system of conditional statements, as stated below.

The diagnosis may be ‘appendicitis’ and management may require an appendectomy when one of the following conditions occurs.

• A male child experiences right lower quadrant abdominal pain, and his WBC is above 20,000/mm³.

• A male child experiences right lower quadrant abdominal pain that lasts from 4 to 24 h, combined with frequent (more than three times) vomiting.

• A male child who already visited the ER in the past 24 h experiences right lower quadrant abdominal pain, combined with frequent (more than three times) vomiting.

• A child experiences right lower quadrant abdominal pain, combined with frequent (more than three times) vomiting, and his or her WBC is above 20,000/mm³.

• A child experiences right lower quadrant abdominal pain, combined with a temperature of 37°C to 39°C, and his or her WBC is above 20,000/mm³.

The diagnostic decision may be ‘resolution’, and management may be discharge when one of the following conditions occurs.

• A child experiences abdominal pain (neither right lower quadrant nor suprapubic) that lasts from 4 to 24 h.

• A child experiences intermittent abdominal pain (neither right lower quadrant nor suprapubic).

• A child experiences abdominal pain (neither right lower quadrant nor suprapubic) that is not accompanied by vomiting.

• A child experiences abdominal pain (neither right lower quadrant nor suprapubic), combined with a normal temperature, and his or her WBC is between 10,000/mm³ and 20,000/mm³.

• A child experiences intermittent, nonlocalized pain, combined with a normal temperature, and his or her WBC is between 10,000/mm³ and 20,000/mm³.

The relative importance of the attributes in the statements that comprise the clinical algorithm is further supported by an analysis of the relative frequencies calculated for the interesting rules during the validation tests (where a frequency of 100% denotes the most frequent attribute) (Figure 1).

Figure 1.

Frequencies of attributes appearing in the interesting rules generated during reclassification tests. An analysis of the attributes in the interesting rules confirms that an effective clinical algorithm for the classification of the child with abdominal pain into one of two decision classes (appendicitis or resolution) must include the following: location of pain (Abd Pain-Site), type of abdominal pain (AbdPainType) and a white blood cell count (WBC). A frequency of 100% is assigned to an attribute appearing most frequently in the interesting rules. A label ‘correct’ refers to those interesting rules that classified patients correctly, while ‘All Rules’ refers to all of the interesting rules, regardless of the correctness of their classification. AbdPain-Dur Duration of abdominal pain; PreVis Previous visits to the emergency room in the past 48 h; Temp Temperature; Vomiting Number of times vomiting occurred

In Figure 1, the label ‘Correct’ refers to those rules that classified patients correctly, while ‘All rules’ refers to all of the interesting rules, regardless of the correctness of their classification. Figure 1 confirms that an effective clinical algorithm must include the following medical information: location of abdominal pain, type of abdominal pain and WBC. The other diagnostic items may provide useful supplementary diagnostic information for the management of appendicitis.

DISCUSSION

The purpose of the present study was to develop a clinical algorithm for the triage of a child with abdominal pain. A set of 12 clinical symptoms and signs (attributes) that are routinely considered in the diagnosis of appendicitis were reduced to nine attributes that were identified for the development of the algorithm.

The clinical algorithm developed reflects a physician’s inductive reasoning when diagnosing abdominal pain. The algorithm primarily evaluates abdominal pain site and type, together with WBC, further supported by information about the patient’s pain duration, fever, vomiting, sex, history of previous visits to the ER and age. The sensitivity of a clinical algorithm in the diagnosis of abdominal pain may be less than that of ultrasonography or computed tomography, but, on the other hand, an algorithm is very easy to implement. Thus, our study suggests that with appropriate methodology, clinical data may be used efficiently in, at least, the triage, if not the final diagnostic decision, without resorting to expensive diagnostic procedures. The clinical algorithm that was developed requires field testing and validation before it is used in a clinical setting. A prospective assessment of the algorithm is being organized at two teaching hospitals in Canada.

Attempts have been made to develop diagnostic scoring systems that support the management of the patient with abdominal pain (12,13). However, the usefulness of these systems relies on the assumption that they are operated by a sophisticated, trained user. This is not required for the clinical algorithm presented in this paper because the algorithm can be understood by nonmedical users and, thus, it can be introduced easily into any health care environment.

CONCLUSIONS

A similar approach to that used during this study may be applied to study data related to other specific complaints that are referred to the paediatric ER. The selection of appropriate clinical symptoms and signs, and the development of a structured clinical algorithm for various conditions, such as chest pain, headaches, joint pains and breathing difficulties, may allow rapid management and the streamlined triage of patients. As a result, it is not unreasonable to envisage satisfied parents, staff that are less stressed and, ultimately, an improvement in the management of paediatric emergencies.