Prior Hospitalization for Epilepsy, Diabetes, and Stroke and Subsequent Glioma and Meningioma Risk

There is strong and consistent evidence from both case-control (1) and cohort (2-4) studies that epilepsy is associated with increased risk of both glioma and meningioma. Schlehofer et al. found an increased risk of glioma associated with a history of epilepsy [risk ratio, 6.55; 95% confidence interval (95% CI), 3.40-12.63] that declined when only those diagnosed with epilepsy lasting at least 20 years were included (1). Lamminpaa et al. also observed increased risks of both meningioma and glioma associated with use of antiepileptic medication (3). However, because glioma risk has been shown to increase with proximity to brain tumor diagnosis and decrease with duration of epilepsy, epilepsy results from, rather than causes, meningioma or glioma (5). Although previous cohort studies have described preclinical risk of brain tumors among epilepsy patients, they have had relatively small numbers of brain tumor patients and have therefore been unable to examine their findings in detail by histologic site or age. Previous brain tumor case-control studies, however, have had many brain tumor patients but have used self-reported medical histories, which may be biased, particularly for high-grade glioma (HGG) respondents, a large proportion of whom are proxy respondents (6).

In contrast to the positive association with epilepsy and glioma, an inverse association between diabetes and glioma has been observed previously in two studies (7, 8) but not in four others (1, 9-11). Schlehofer et al. found that stroke increases glioma risk (12). In a later report, Schlehofer et al. also saw an increased risk of glioma among people reporting a history of meningitis (1). Finally, Wrensch et al. observed higher levels of antibodies to herpes simplex virus among glioblastoma patients than among controls (13). Because this virus can cause encephalitis, we also examined the association between encephalitis and brain tumor risk.

Most of the literature on medical conditions and brain tumors assumes that medical conditions, except possibly seizures or epilepsy, precede brain tumor development and play a role in brain tumor etiology (1, 6, 8, 14, 15). However, if epilepsy results from, rather than causes, brain tumors, it may not be the only chronic disease that behaves in this fashion. Therefore, in addition to evaluating associations between diabetes, stroke, meningitis, and viral encephalitis and brain tumors, we examined their temporal patterns relative to time of brain tumor diagnosis to see if inferences can be made as to whether these conditions precede or follow brain tumor development. Medical conditions that result from the preclinical presence of brain tumors will suggest the length of the preclinical period during which brain tumors may be present and therefore possibly susceptible to intervention.

We first identified cases of glioma and meningioma, among individuals who were ages ≤15 years, reported to the Swedish Cancer Registry from 1987 to 1999 [1,501 low-grade glioma (LGG), 4,587 HGG, and 4,193 meningioma]. We included the young adult group in our analysis (ages 15-19 years) because this group is in transition between childhood and adult primary brain tumor patterns. Thus, classifying brain tumor patients ages 15 to 19 years at diagnosis either as children or as adults is somewhat arbitrary. However, we focus our analysis on histologic types of tumors prevalent during adulthood (16).

The Swedish Cancer Registry classifies tumors according to the International Classification of Diseases (ICD)-7 in combination with pathology codes. In addition, from 1993 onward, they used ICD-10 together with histologic classifications from ICD for Oncology-2 (16). We included brain tumor cases in the present study if their ICD-7 codes were 193 and their pathologic codes were 461 (meningioma), 475 (LGG), or 476 (HGG). After 1993, cases were coded using both ICD-7 and ICD-10. Using the ICD-10 code (C71), we were able to identify glioblastoma (ICD for Oncology-2, 9440-9442 and 9460) and anaplastic astrocytoma (9380 and 9401/9403) separately. Therefore, to include all reported glioma cases between 1987 and 1999 in our analysis, we retain the glioma categories low and high grades. However, when there are a sufficient number of glioblastoma and anaplastic astrocytoma cases, we analyze these tumor types separately.

Controls for the present study were randomly selected from the continuously updated Swedish Population Registry from among individuals never diagnosed with cancers of the brain (or, in addition, the pancreas), lymphoma, or leukemia (controls were selected simultaneously for studies of the latter three tumors in addition to the present study of brain tumors). On December 31 of each year of the study, controls were frequency matched to cases for that year on age and sex. Therefore, the year of diagnosis of cancer cases to which controls were matched determined their reference year. The number of controls selected depended on cases' ages at diagnosis. For each cancer case ages >19 years at diagnosis, two controls were selected. For cases diagnosed at ages 15 to 19 years, five controls were selected. Thus, controls represent the population that yielded the cases (except, of course, that control year, age, and sex distributions are more like those of cases than they would have been had they not been matched to cases).

Next, case-control data were linked to the Swedish Hospital Discharge Registry from 1969 to 1999 where discharge diagnoses of epilepsy, diabetes, stroke, allergy, asthma, autoimmune diseases, and infections reported previously to be associated with brain tumor were identified (1, 6, 8, 14, 15). We restricted the analysis to discharge diagnoses before the date of brain tumor diagnosis. From 1969 to 1996, as many as six discharge diagnoses were reported for each hospital discharge, and from 1997 to 1999, as many as eight were reported. We used all available discharge diagnoses in the analysis, making no distinction between primary and other diagnoses. We could not examine differences between type I and II diabetes because information on diabetic type was not included in Swedish Hospital Registry data until 1997.

To assure adequate statistical power, we confined our study to the most common primary adult brain tumors: LGG, HGG, and meningioma. In addition, we restricted our analysis of previously identified medical conditions to discharge diagnoses found in at least five future LGG, HGG, or meningioma patients during one of the periods on which we focused (e.g., Table 2).

We used unconditional logistic regression to compare case and control odds of discharge diagnoses and to adjust for age, sex, and year of diagnosis or reference year. Results for epilepsy did not vary by sex; therefore, we do not present epilepsy findings for males and females separately. Each odds ratio (OR) in Tables 2, 3, 5, and 6 represents a case-control comparison of disease odds during the preclinical period indicated. We do not report summary odds ratios because these time-related variables modify the effects of associations between epilepsy, diabetes, or stroke and brain tumor risk (17).

Table 4 is based on a different sample from the other tables because it includes only brain tumor cases and controls diagnosed with epilepsy. Brain tumor cases and controls not diagnosed with epilepsy could not be included in Table 4 because we do not have a variable representing their age at epilepsy diagnosis. Here, the reference group consists of cases and controls diagnosed with epilepsy at age 45 years.

Trend tests for epilepsy are reported in footnotes to Table 2. Although our hypothesis that prediagnostic epilepsy reflects presence of a brain tumor suggests that year of brain tumor diagnosis or reference year should be included with the prediagnostic period, results of such trend tests may be unduly influenced by large ORs for year of brain tumor diagnosis. We therefore also present trend tests excluding year of diagnosis or reference year. These statistical findings should be evaluated together with inspection of the ORs on which they are based.

Because we were interested in the earliest time of occurrence of each of the five medical conditions that we investigated, we included only the first diagnosis for each condition during the observation period (1969-1999). Once a case or control was diagnosed with a particular condition, they were excluded from subsequent analyses of that condition because they were no longer part of the population at risk. There were some patients who were discharged with more than one of the conditions in which we were interested. Because epilepsy is strongly associated with brain tumor risk (14), we included all first time epilepsy diagnoses in Tables 2, 3, and 4 whether they were accompanied by stroke or diabetes. However, because we were interested in the independent relationship between diabetes and stroke and brain tumors, we excluded all patients from the analyses of diabetes and stroke (Table 5) if they had also been diagnosed with epilepsy or both diabetes and stroke during the same hospital stay (see Table 1).

Two major differences among demographic variable distributions can be seen in Table 1. First, controls were matched on age and sex to tumors at sites other than the brain and are therefore, on the average, older than cases and have a different sex distribution as well (Table 1; see Materials and Methods). Second, glioblastoma multiforme and anaplastic astrocytoma were not classified separately until 1993 and therefore have higher median years of diagnosis than do other types of brain tumors (see Materials and Methods). These differences provide the rationale for adjusting ORs in subsequent analyses for age, sex, and year of diagnosis.

Also shown in Table 1 are percentages of cases and controls with discharge diagnoses of epilepsy, diabetes, and stroke or a combination of these diseases. Numbers adjacent to percentages include the total as well as the number remaining after excluding patients with more than one of these discharge diagnoses during the same hospital stay (see Materials and Methods). Because the association between each of these three discharge diagnoses and brain tumor risk is modified by time to brain tumor diagnosis or age at brain tumor diagnosis (Tables 2,Table 3,Table 4,Table 5,-6), no summary ORs are shown in Table 1.

LGG risk is elevated among patients discharged with epilepsy ≥11 years before diagnosis and peaks at the year of diagnosis (Table 2; P for trend excluding year of diagnosis = 0.0001). An elevated risk of HGG occurs almost as long before diagnosis; however, preclinical ORs are generally lower and the preclinical increase in glioma risk is weaker (P for trend excluding year of diagnosis < 0.0001). ORs for glioblastoma multiforme and anaplastic astrocytoma follow a pattern similar to those for HGG but start increasing slightly closer to the time of diagnosis (P for trend tests = 0.0003 and 0.1598, respectively). In contrast, the association between epilepsy and subsequent meningioma diagnosis increases only slightly until year of diagnosis (P for trend excluding year of diagnosis = 0.4953 and P for trend including year of diagnosis < 0.0001).

In Table 3, a strong epilepsy-age at brain tumor diagnosis interaction is shown for LGG and HGG (P < 0.01 for age-epilepsy statistically significant interaction for all three preclinical periods), but for meningioma the interaction is restricted to the year of meningioma diagnosis (P < 0.01). The last three rows of Table 3 indicate that prevalence of epilepsy among controls as well as among cases contributes to variation of ORs by age at diagnosis. Specifically, there are relatively fewer controls diagnosed with epilepsy in the youngest age category (<55 years) and slightly more controls diagnosed in the oldest age category (≥75 years).

In addition to being modified by age at brain tumor diagnosis (Table 3), the association between epilepsy and brain tumor risk also varies with age at first epilepsy diagnosis (Table 4). A pattern similar to Table 3 is seen for LGG and HGG and meningioma ≥4 years before diagnosis; however, within 4 years of HGG and meningioma, ORs form a plateau across ages ≥45 years.

HGG ORs for diabetes and stroke (Table 5) differ from those seen for epilepsy (Table 2). They vary little until the year of diagnosis when they change direction and increase (Table 5). [Average diabetes ORs (95% CIs) for HGG over the period excluding year of diagnosis are 0.65 (0.49-0.86) for males and 0.63 (0.44-0.90) for females.] Diabetes ORs (95% CIs) for glioblastoma multiforme for both sexes combined are similar to those for HGG as a whole [≥4 years before diagnosis, 0.52 (0.29-0.92); 2-3 years before diagnosis, 0.90 (0.40-2.02); and year of diagnosis, 10.94 (7.11-16.82)]. Stroke ORs for HGG essentially follow the same pattern as do those for diabetes. [Average stroke ORs (95% CIs) excluding year of diagnosis are 0.68 (0.51-0.91) for males and 0.54 (0.35-0.83) for females.] Again, glioblastoma multiforme ORs (95% CI) for stroke are similar to those for HGG [≥4 years before diagnosis, 0.52 (0.28-0.98); 2-3 years before diagnosis, 0.30 (0.10-0.95); and year of diagnosis 12.94 (9.86-16.61)].

Male meningioma ORs for both diabetes and stroke increase beginning 6 to 7 years before brain tumor diagnosis. However, female meningioma diabetes ORs are more difficult to characterize because they initially decline (P for trend = 0.02) until 1 to 2 years before meningioma diagnosis and subsequently increase (P for trend = 0.0001). Female meningioma stroke ORs are similar to those for males.

Numbers of observations for LGG are too few to analyze by observation periods shown in Table 5, but average ORs (95% CIs) for both diabetes and stroke are close to the null [excluding year of diagnosis, 0.73 (0.40-1.33) for males and 1.02 (0.54-1.92) for females], whereas those for stroke are similarly close to the null [excluding year of diagnosis, 1.19 (0.68-2.07) for males and 1.06 (0.50-2.26) for females].

We observe an age at brain tumor diagnosis-stroke interaction for all three tumor types (Table 6). In general, the younger age group (<65 years) has the highest risk of stroke; however, this is due in part to the smaller proportion of controls discharged with stroke among younger people. Finally, meningitis and viral encephalitis are associated with increased brain tumor risk near the time of brain tumor diagnosis but not before (Table 7).

Although meningioma and LGG are characterized as slow-growing tumors with long preclinical phases (18) and HGGs as fast-growing tumors with relatively brief preclinical phases (19), our findings for preclinical epilepsy suggest unexpected similarities in the length of the preclinical phase of these tumor types (Table 2). In contrast, results for diabetes and stroke are more consistent with the putative dichotomy between slow-growing and fast-growing tumors. Diabetes and stroke ORs for meningioma suggest a slight increase in risk several years before diagnosis, whereas diabetes and stroke ORs for HGG do not (Table 5). ORs for meningitis and viral encephalitis present a different pattern; their proximity to time of tumor diagnosis indicates that these infections are involved only in the late stages of brain tumor development (Table 7).

These varying patterns of preclinical disease may indicate the following: although some gliomas begin to develop several years before diagnosis, their final rapid growth, producing systemic effects, occurs late in their development. In contrast, meningiomas produce systemic effects earlier as shown by altered risk of meningioma among people discharged with diabetes and stroke several years before diagnosis. Our results warrant further research to identify preclinical changes, in addition to seizures or epilepsy, that may eventually allow early intervention in meningioma and even glioma development.

Not all regions of Sweden were served by the Hospital Discharge Registry in 1969, the first year that we begin using Registry Data. By 1980, however, 80% of Swedish regions were covered by the Registry, and after 1987, the Registry served the entire country. There are no differences in brain tumor incidence rates by region in Sweden so that failure to register hospital stays would not be expected to differ between cases and controls and would therefore result in nondifferential misclassification of discharge diagnoses. Thus, reported ORs may be biased toward their null values.

Our findings depend on the quality of Swedish Hospital Registry data, although nondifferential error would probably not affect internal validity. Nilsson et al. examined clinical records from hospital stays of 995 patients whose discharge diagnoses were reported to the Swedish Hospital Registry. They found that 17.3% of the discharge diagnoses recorded in the Hospital Registry had an error in the primary diagnosis. These percentages, however, are based on coding on a five-digit level (20). Discharge diagnoses used in the present study would have a somewhat lower error rate because they were based on a four-digit level code (with broader diagnostic categories than the five-digit code) and included as many as eight discharge diagnoses for each patient.

When interpreting results of the present study, it is important to note that individuals with discharge diagnoses of epilepsy, diabetes, or stroke differ from those who have these chronic diseases but are not hospitalized. For example, those hospitalized may have more severe illness in general or, more specifically, have failed to control their diabetes. By including as many as eight discharge codes (not just the primary discharge diagnosis), however, our discharge diagnosis groups may be somewhat more representative of nonhospitalized patients than they would have been had we studied primary discharge diagnoses only, in that they may include some patients who were not initially hospitalized for these three conditions but were initially hospitalized for other reasons.

Our results confirm findings from two cohort studies of people with seizure disorders or epilepsy (2, 4) and one cohort study of patients using antiepileptic medication (3). In all three cohorts, risk of brain tumors increases with proximity of epilepsy to time of tumor diagnosis and decreases as the time between epilepsy and brain tumor diagnosis increases. The implications of this finding are that if epilepsy medication causes brain tumors then medication used for a short duration would be associated with a greater brain tumor risk than would medication used for a long duration. Such a duration-response relationship would be improbable.

Further evidence that epilepsy occurs early in the process of tumor development is provided by the observation that risk of LGG is greater among people discharged with epilepsy than is risk of HGG (Table 2). Although this finding may suggest that epilepsy reflects different causal mechanisms in the two tumor types, it is also consistent with epilepsy flagging tumors earlier in their development for diagnosis. In addition, epilepsy discharge diagnoses are associated with a greater risk of brain tumors among people diagnosed with brain tumors at younger rather than older ages (Table 3). Also suggesting that epilepsy causes diagnosis at an earlier stage of brain tumor development is the fact that a history of previous seizures is associated with a better prognosis for both LGG and HGG patients (21, 22).

A study of comparative brain locations of LGG and glioblastoma multiforme conducted by Duffau and Capelle findings provides some support for the suggestion that epilepsy itself may cause early diagnosis of low-grade tumors (23). They found that lower-grade tumors tend to be located closer to eloquent regions of the brain than do higher-grade glioblastomas. (Eloquent brain regions may be defined as functionally important regions that would result in neurologic deficits if injured.) This proximity of low-grade tumors to eloquent regions may account for their earlier production of symptoms, perhaps including epilepsy, leading to earlier diagnosis. Nonetheless, some people eventually diagnosed with high-grade tumors are diagnosed with epilepsy a relatively long time before their tumor is identified, suggesting that high-grade tumors have a longer prediagnostic period than is generally recognized.

Little is known about the structural, biochemical, or histologic mechanisms that alter the peritumoral environment to produce seizures (24); however, HGGs seem to be able to produce these changes before they can be seen on magnetic resonance imagery scans and therefore before their rapid growth period (25). If HGGs can produce seizures at an early stage in their development, our findings of a relatively long preclinical period for HGG and even glioblastoma multiforme may be consistent with what is known about the rapid rate of glioblastoma growth possibly during a later stage immediately preceding diagnosis.

The distinction between Tables 3 and 4 is important. In Table 3, the modifying variable is age at brain tumor diagnosis. For glioma and meningioma during the year of brain tumor diagnosis, the association between epilepsy and brain tumor risk is strongest among people who are ages <55 years at brain tumor diagnosis. This finding results in part from the greater risk of epilepsy at older than younger ages (last three rows of Table 5) but also indicates that epilepsy may occur at an earlier stage of brain tumor development (producing a younger brain tumor patient at diagnosis) than other symptoms that eventually lead to brain tumor diagnosis.

In Table 4, however, the age variable is age at epilepsy diagnosis. In this analysis, younger people discharged with epilepsy seem to be at greater risk of a brain tumor when the time of brain tumor diagnosis is at least 4 years in the future. However, within 4 years of brain tumor diagnosis, associations between epilepsy and HGG or meningioma risk change. Now, risk of brain tumors is greater for people diagnosed with epilepsy at ages ≥45 years than it is for people who are diagnosed when they are ages <45 years. This finding reflects the fact that epilepsy is a common presenting symptom and therefore follows the age distribution of patients at time of HGG or meningioma diagnosis. Thus, Table 4 reflects the fact that epilepsy can either lead to relatively rapid brain tumor diagnosis (the most common case in our data) or occur several years before brain tumor diagnosis. The message for the clinician from Table 4 (who obviously does not know whether the newly diagnosed epilepsy patient has a brain tumor or not) is that a new diagnosis of epilepsy in an adult at any age may indicate the presence of HGG or meningioma. Younger age at diagnosis of epilepsy, however, is associated with a greater risk of LGG than is an older age at epilepsy diagnosis.

The inverse association seen during the preclinical period between diabetes and glioma (Table 5) has not been observed consistently (1, 7-11). This lack of consistency may be attributable in part to varying times of ascertainment of medical conditions relative to time of brain tumor diagnosis. Although we find that diabetes is associated with reduced risk of HGG, we find no duration-response effect. Perhaps we do not observe this effect because, once identified, diabetes can be treated and possibly consequently lose its ability to afford protection against brain tumors. We were unable to replicate Brenner et al.'s finding of an age-diabetes interaction; however, they measured a history of self-reported diabetes rather than hospitalization discharge diagnosis of diabetes, so their diabetes variable differs from ours (8).

We cannot compare our tumor type–specific and time to diagnosis–specific findings for preclinical stroke with those of Schlehofer et al. who reported a combined OR of 1.9 for self-reported stroke occurring ≥2 years before diagnosis of either meningioma or glioma (12) In the present study, the association between stroke and HGG did not vary with the length of time to brain tumor diagnosis until the year of brain tumor diagnosis. It is possible that stroke, like diabetes, reduces risk of HGG only when the stroke initially occurs. The beneficial effect of stroke may come from their ability to reduce the blood supply to a potential tumor. This hypothesis could be evaluated by comparing the relative location in the brain of strokes occurring before brain tumor diagnosis among cases and controls and also, among cases, comparing the location of the stroke with that of subsequent HGGs. The increased risk of brain tumors among patients diagnosed with stroke during the year of brain tumor diagnosis probably represents either misdiagnosis of brain tumors as strokes (26) or physiologic responses to tumor growth (27). In either case, the increase during the year of diagnosis is probably attributable to the presence of the tumor.

Our findings for meningitis are similar to those reported by Schlehofer et al. (1) [OR (95% CI) for glioma 1.22 (0.69-2.17) and 1.28 (0.54-3.03); both ORs are based on a self-reported history ≥2 years before brain tumor diagnosis]. There is an extensive but largely hypothetical literature on infectious disease as a cause of primary adult brain tumors (14); however, because of the proximity of the increased risk of meningitis and viral encephalitis to brain tumor diagnosis, we conclude that these diseases probably result from, rather than cause, brain tumors.

Since the late 1970s, computed tomography scans have been available in Sweden and have been used to identify brain tumors among patients diagnosed with a first adult episode of epilepsy of unexplained origin (e.g., drug or alcohol withdrawal and head trauma would be explained by the events with which they are associated). Magnetic resonance imaging became available throughout Sweden in the 1990s and also sometimes used to diagnose brain tumors. When making inferences about the length of the preclinical period from our findings to countries other than Sweden, it is important to remember that there is great international variability in access to health care. Therefore, even if the customary standard of care dictates that neuroimaging studies be conducted on all first time adult epilepsy patients, actual practice depends, in addition to medical policy, on the distribution of health care resources.

The importance of our findings extends beyond the small number of exposed cases that we observed because only a small percentage of people with epilepsy, diabetes, and stroke are hospitalized. Previous brain tumor research on people with these diseases who were not hospitalized suggests that our findings have applicability to all people with these diagnoses whether they are hospitalized (14). Subsequent research may be able to determine whether the length of the preclinical period among people without epilepsy is similar to that observed for people with epilepsy. That is, we need to find out whether epilepsy is a marker for a slow-growing tumor or whether it merely occurs in tumors of typical preclinical length causing early diagnosis. Evidence for long-term preclinical physiologic changes, in addition to epilepsy, should be sought. If such evidence is found it could, in the future, be combined with neuroimaging techniques for earlier identification of cases. Research on additional preclinical medical conditions that alter brain tumor risk would be a place to begin this research. Further studies of cohorts with serial biomarker measurements would also be important. Of particular interest, because of the observed inverse associations with diabetes and stroke, is the preclinical brain tumor risk among people diagnosed with other cardiovascular diseases.