A Data Mining and Survey Study on Diseases Associated with Paraesophageal Hernia

Abstract

Paraesophageal hernia is a severe form of hiatal hernia, characterized by the upward dislocation of the gastric fundus into the thoracic cavity. In this study, the 1999 National Inpatient Sample dataset of the Healthcare Cost and Utilization Project was analyzed using data mining techniques to explore disorders associated with paraesophageal hernia. The result of this data mining process was compared with a subsequent expert knowledge survey of 97 gastrointestinal tract surgeons. This two-step analysis showed that the results of data mining and expert knowledge are consistent in some factors that are highly associated with paraesophageal hernia: older age, other gastrointestinal tract disorders and obesity, for example. But the data mining approach revealed some other related disorders that were not known to the experts or reported in the literature, for example, hypertension, peritoneal adhesions and gall bladder/bile duct diseases. These findings lay a framework for subsequent hypothesis-driven research.

Introduction

A hiatal hernia is a herniation of part of the stomach into the thoracic cavity through the esophageal hiatus of the diaphragm. The existence of this disease has been acknowledged for more than 400 years. It is recently getting attention because of its association with gastroesophageal reflux disease [1–4] and because of the life-threatening condition caused by strangulation and perforation of paraesophageal hernias (PEH), a severe form of hiatal hernia (Fig. 1). Hiatal hernias may be classified into three types. The first type is the “sliding” (type I) hiatal hernia, where the esophagogastric junction (EG junction) is above the diaphragmatic hiatus. Figure 1-B shows the upward dislocation of the EG junction and may be compared with the normal anatomic structures as shown in Figure 1-A. Type I is by far the most common type of hiatal hernia, with seven times higher incidence rate than the other two types [5]. The second type is the “paraesophageal” (type II) hiatal hernia, where the EG junction is at the normal position but the widened hiatus allows the fundus of the stomach to protrude into the chest (Figure 1-C). This type is very rare. Mixed hiatal hernias (type III hernias) are a little more common than type II and are characterized by both an upward dislocation of the EG junction and by protrusion of the gastric fundus into the chest.

Figure 1. The anatomy of Paraesophageal Hernias (PEH).

A. Normal anatomic structure at the esophagogastric junction. B. sliding hiatal hernia. C. paraesophageal hernia. Reprinted from [1] with permission from Excerpta Medica Inc.

Although types II and III hernias are both rare, they are far more dangerous than type I hernias: 25% of patients with these two types of hernia develop incarceration and strangulation of the hernia [5, 6]. If these emergency conditions occur, the mortality rate is high [5, 6]. Early detection with evaluation and possible surgical management of PEH (for convenience, both type II and III will be called PEH for the remainder of this document) is the key to avoiding these life-threatening complications.

Exploratory studies of PEH disease association can be used to identify targets for further hypothesis-driven research on causal relations. In this study, both an exploratory approach (data mining) and an expert survey were used to explore the disorders possibly associated with PEH. The results are compared and the advantage of each method is highlighted.

Methods

Data Source

The data used for this study was the 1999 Nationwide Inpatient Sample (NIS) database. NIS is part of the Healthcare Cost and Utilization Project (HCUP). The NIS contains the patient-level clinical and resource use information included in a typical discharge abstract [7]. It is designed to approximate a 20-percent stratified sample of U.S. community hospitals. All discharges from the sampled hospitals are included in the NIS database. There is one dataset for each year from 1988 to 2003, each containing from five to eight million stays. The information for each stay includes primary and secondary diagnoses and procedures, patient demographics, hospital characteristics, and charge information.

Data Preprocessing

SAS® for Windows version 8 was used for data preprocessing. Irrelevant data elements were first excluded leaving the following variables: age, primary and up to 14 secondary diagnoses by ICD-9-CM codes, primary and up to 14 secondary procedures by ICD-9-CM codes, gender, state postal code for the hospital, expected primary payer, race and median household income for patient’s ZIP code.

After consultation with GI surgeons and coding experts at this institution, cases with PEH were selected using the ICD-9-CM procedure code 53.7 (which corresponds to the more commonly used CPT code of 39502: Repair, paraesophageal hiatus hernia, transabdominal). Other procedural codes were felt to be less specific and no diagnostic codes are available that clearly identify PEH patients. Since PEH is almost always treated with surgery, it is unlikely that we have excluded a significant number of cases with this criterion. A total of 1633 PEH cases were found. Note that even though the dataset is a stratified sample, we took all PEH cases without stratification. The justification is that PEH is a rare disease (1633 out of 8 million cases in this dataset) and we do not expect it to follow a normal distribution.

In order to exclude congenital hiatal hernias, which are likely to be etiologically different from adult hiatal hernias, only patients aged 18 years or older were selected for both case sets.

Association Rule Mining

Association rule mining discovers the association relationships among a set of items, i.e. detecting items that frequently occur together in a transaction. Here, one transaction is one inpatient stay with a set of diagnosis codes. An association rule found in the mining process might look like this: A ⇒ B, i.e. if A is present (A being one or more item in the set) then, with a selected measure of strength, B is likely to present (B being one or more item in the same set). Measures of the strength of the association are support and confidence. Support is the percent of transactions that contain both A and B (how likely AB co-occur in the whole dataset), and confidence is the percent of transactions containing A which also contain B (how likely is B in the set of transactions that has A). The higher these two measures are, the stronger the association.

Because of the transactional nature of the data, association rule mining was used to discover diseases that are associated with PEH. Since we filtered the dataset for transactions with PEH, we used support as our measure. It was performed with Classification Based on Associations (CBA, School of Computing, National University of Singapore) [8, 9]. The set of 15 diagnosis codes in the 1633 PEH cases were used as a transaction set, with each of the codes an item in that set. A minimum support of 0.6% and minimum confidence of 50% were chosen. Since this is the hypothesis-generating step, the low 0.6% support was deliberately chosen in hopes of finding associated diseases even though they have a low prevalence.

The associated diagnoses discovered in this process were analyzed to remove acute conditions caused by surgery, since all PEH cases had surgery. For instance, ‘pulmonary collapse’ was removed because it is very likely to be related to the surgery for PEH rather than to the disease itself. The remaining diagnoses were grouped into disease groups. The grouping was done empirically by judging a number of parameters, including the diseases’ similarity, probability that the codes represented the same disease, and likelihood that codes represent diseases having the same cause. The 40 disease groups generated in this process were used as variables in the survey of experts.

Expert Knowledge Survey

The 40 associated diagnostic groups discovered in the association rule step, together with some demographic variables of the dataset such as age, race, expected primary payer (private, Medicare/Medicaid or self pay) and median household income for patient’s ZIP code were assembled in a web survey form which asked respondents to indicate all factors or conditions that they suspected to be associated with the development of paraesophageal hernias. Electronic mail invitations were sent to 900 members of the Society for Surgery of the Alimentary Tract (SSAT), randomly selected from SSAT’s 2004 roster. Electronic reminders were sent to members who had not completed the survey after three weeks. Three months after the original solicitation, the responses were collected and analyzed. This survey was approved by the Institutional Review Board of Oregon Health & Science University, IRB #8412.

Results

The 1999 NIS dataset contains 1633 cases identified as having PEH using the ICD-9-CM procedure code 53.7. Of these, 1043 (63.9%) are female and 91.1 % are white (of the 1056 cases where race was indicated). Mean age is 60.5 years with a standard deviation of 17.2 years. Over 94% had ZIP codes where average household income was higher than $25,000/year.

Association Rule Mining

The result of the association rule mining is a list of diseases derived from the left hand side of the association rules with presence of PEH being the right hand side. The disease with the highest support (38.9%) was esophageal reflux disease (ICD-9-CM 530.81), which is not surprising since the connection between this and PEH has been well documented [1–4]. Table 1 shows the 10 diseases with the highest support and the diagnostic groups to which they were assigned. Acute conditions related to surgery are not included. A total of 89 diagnoses leading to 40 diagnostic groups were discovered.

Table 1. The ten diseases with highest support for association with PEH, based on association rule mining.

The diagnostic groups chosen for further analysis based on these diseases and the ICD-9-CM codes for those groups are also listed. Note that other diseases with lower support may also contribute to the choice of diagnostic groups.

ICD-9-CM	Description	Support	Diagnostic Group (ICD-9-CM diagnostic codes)	Support for diagnostic group
530	Esophageal reflux	38.9	Other Upper GI diseases (530–537,787.1,787.2)	71.2
401.9	Hypertension nos	28.4	Hypertension (401)	30.5
278.01	Morbid obesity	13.3	Obesity (278)	15.6
568.0	Peritoneal adhesions	8.6	Peritoneal adhesions (568.0)	8.6
574.10	Cholelith w cholecys nec	6.9	Gall bladder/bile duct diseases (574–576)	13.8
427.31	Atrial fibrillation	6.2	Dysrhythmia (427, V450, V4589)	9.7
244.9	Hypothyroidism nos	6.0	Hypothyroidism (244)	6.0
496	Chr airway obstruct nec	5.9	Chronic respiratory diseases (490–496,786.09)	15.2
272.0	Pure hypercholesterolemia	5.0	Hyperlipidemia (272)	6.4
250.00	DMII wo cmp nt st uncntr	5.0	Diabetes (250)	5.9

Expert Knowledge Survey

Out of the 900 electronic invitations sent, 236 were undeliverable. Of the 664 members assumed to have received the invitations, 97 (15%) completed the survey. Table 2 shows, for each diagnostic group and demographic factor, the count and percent of respondents who indicated that the diagnostic group or factor is suspected to be associated with development of PEH. The result for some demographic factors was very consistent with the findings from the dataset, i.e. PEH appears to happen in an older and white population. Less than half (40/97) of the respondents noted an association of household income with PEH. The respondents' acknowledgement of Medicare/Medicaid as an association is correlated with older age.

Table 2. Results of the expert knowledge survey determining demographic factors or medical conditions suspected to be associated with the development of PEH.

A. Demographic factors in the survey and the count (percentage) of participants who selected that factor. B. Medical conditions chosen most often as having an association with the development of PEH, from 40 condition groups.

A
Demographic Factors	Count (Percentage)
Age
Younger Age	4 (4.1%)
Older Age	93 (95.9%)
Race
White	48 (49.5%)
African American	2 (2.0%)
Asian	2 (2.0%)
Other	0 (0.0%)
Household income level
>45K	9 (9.3%)
35K–45K	13 (13.4%)
25K–35K	9 (9.3%)
<25K	9 (9.3%)
Insurance Status
Private insurance	5 (5.2%)
Medicare/Medicaid	27 (27.8%)
Self-pay	2 (2.0%)
B
Medical Conditions	Count
Obesity	68
Other Upper GI diseases	55
Other hernias (umbilical,	31
Chronic respiratory	29
Tobacco use	21
Collagen vascular diseases	21
Diabetes	14
Sleep apnea	12
Female stress incontinence	11
Anemia	10
Malnutrition	10
Osteoporosis	10
Colonic diverticulae	10

Four medical conditions -- other GI diagnoses, obesity, respiratory diseases and diabetes -- were in the top lists of medical conditions found both in the survey and by the association mining algorithm. Other conditions chosen frequently in the survey included existence of other hernias, tobacco use and collagen vascular diseases, which did not have high support from data mining. Conversely, hypertension, peritoneal adhesions, gall bladder/bile duct diseases and other conditions with higher levels of support by data mining were not supported by expert opinion.

Discussion

In this study a large dataset was analyzed using exploratory data analysis techniques and the result was compared with results of a survey of GI surgeons’ expert knowledge. This two-step analysis is able to show that data mining technology can be used effectively as a preliminary study methodology for exploring the data and finding possible predictors where no prior experience or scientific theories are available to aid the analysis. The data mining study uncovered a number of possible important associated diseases, which may deserve future hypothesis-oriented studies. It is easy to understand the associations in some cases and both approaches are consistent in these aspects. For example, PEH is, in fact, indicated to be associated with other GI diagnoses. The relationship to PEH of GI diseases such as esophageal reflux has been the interest of recent studies [2–4]. It is also not surprising that obesity is associated with PEH, since there are reports in the literature that obesity is associated with gastroesophageal reflux disease [10,11], which again is associated with hiatal hernias. The finding of these known associated factors and its consistence with the finding of the survey assure that this algorithm is working correctly.

Note that some medical conditions appear in both lists of top associated diseases from the two approaches, but some conditions appear only in one or the other. Table 1 suggests associations that may be important if not easy to explain, including the associations with peritoneal adhesions, gall bladder/bile duct diseases, and dysrhythmias. These associations are not known to the experts. This is the value of the data mining approach, i.e. discovery of factors not so apparent to human intuition or experience that may be worthy of further studies.

Some of the disease associations discovered in this process may be confounders rather than true associations. For instance, hypertension has a high support value (28.4), but may be a confounder of the elderly age of PEH cases.

In the survey, 32% and 22% surgeons indicated that development of PEH is suspected to be associated with other hernias and collagen vascular diseases, respectively. Since collagen is the major stress-bearing component of connective tissue, it is reasonable to suspect collagen defects are responsible for the connective tissue weakness that can lead to various kinds of hernias. There are reports in the literature indicating there might be some relationship between the two [12, 13]. There is no direct epidemiological evidence, however. The authors have tested for both potential relations in this NIS dataset and could not reject the null hypothesis at significance level of 0.05 (tests and results were not shown in this paper) in either case.

These discrepancies between the two results highlight the value of using data mining as an exploratory hypothesis-generating tool. Hypotheses generated based on personal experience or even by survey of experts may be prone to bias. For example, a low survey response rate is a problem in that results may be influenced by respondent bias. On the other hand, the publicly available datasets can be a fast, easily accessible solution for preliminary studies, even though there are some issues with the quality of data, as will be discussed.

There are limitations in this study and the first limiting factor is the data itself. The data used in this study were collected from the discharge abstracts of the sampled hospitals, collected mainly for use in hospital facility utilization studies. Using data for purposes other than that for which it was collected may lead to some problems. For example, in the NIS dataset, it is likely that only diagnoses related to the current hospital stay are reported. This leads to the probability that, for patients with PEH, other gastrointestinal diagnoses are more likely to be reported than diagnoses related to other organ systems. In addition, these codes are used for billing purposes rather than to reflect accurately the state of the patient during the hospital stay.

Second, the classification principles of ICD-9-CM codes are not aligned with the needs of this study. For example, we were unable to discern different types of hiatal hernia using the diagnosis codes in the dataset, which made it difficult to identify clear-cut cases of PEH and sliding hiatal hernia. We have substituted procedure codes, but this means that any patients with the diagnosis but who did not have surgery on that hospitalization will not have been detected.

Third, the disease list used in the survey was derived from the data-mining step. Even though we used a low cutoff point, there might be bias in data collection so some diseases of interest were left out.

Finally, our survey had a 15% response rate, which was low but not-unexpected in surgeons. Most of those responses were from clinicians associated with academic institutes. We did not attempt to compare respondents with nonrespondents which may have resulted in responder bias.

Conclusions

In this study, data mining algorithms were used to explore the possible associations between PEH and other diseases or conditions. A comparison was made to a survey capturing GI surgeons’ expert knowledge. The highest associated diseases, based on association rules, are other GI disorders, obesity, hypertension, gall bladder/bile duct diseases and peritoneal adhesions. Further hypothesis-oriented research may be able to elucidate the mechanisms of these associations. It is not surprising to see the associations with obesity and other GI diseases as they are consistent with expert knowledge, and we are especially interested in disorders that are not known to experts. This preliminary study can serve as the starting point for future hypothesis-driven clinical studies. Since data mining is an exploratory and hypothesis-generating process, it can be used to uncover possible associations that are not known or intuitively apparent.