- Top of page
Summary: Purpose: To investigate the extent and causes of the differences in mortality found in studies on mortality in epilepsy based on a quantitative review of the literature.
Methods: We used MEDLINE database and Cumulative Index Medicus for 1960–2001, Excerpta Medica for 1948–1965, and relevant journals and bibliographies. We selected comparative studies investigating mortality in epilepsy patients conducted in the last 100 years. The Standardized Mortality Ratio (SMR) was selected as primary outcome. Nineteen studies were included. Pooled estimates were precision weighted and tested for homogeneity. Sources of variability between risk estimates were explored by using multivariate fixed-effects models.
Results: SMRs ranged from 1.3 to 9.3. Risk estimates proved heterogeneous (χ2 test statistic: 1,177; df = 18; p < 0.001). The most important determinant was “source population,” explaining half of the variance of the estimates (R2, 0.47; p = 0.006). SMRs in community studies ranged from 1.3 to 3.1, and for institutionalized populations, from 1.9 to 5.1.
Conclusions: Our results show that the mortality risk in patients with epilepsy is dependent on source population of patients. Within the different source populations, considerable unexplained variance remains. Hence no uniform summary estimate for the elevated mortality could be determined.
Mortality in patients with epilepsy is reported to be considerably higher than that in the general population (1–4). The excess mortality risk is seen in all age groups, and is generally reported to be higher in men than in women. Epilepsy mortality studies conducted in the last hundred years all demonstrated this elevated mortality, but risk ratios varied considerably (3,4). The extent and causes of this elevation remain unclear, and many factors, such as patient selection, calendar period, and duration of follow-up all possibly contribute in some way to the variation between the estimates. To explain variability between estimates, we carried out a meta-analysis of mortality studies in patients with epilepsy, by reviewing follow-up studies published since 1900.
- Top of page
A total of 245 publications was identified, of which 63 studies specifically studied epilepsy mortality. Of these studies, 17 dealt with epilepsy as a cause of death in the general population and were excluded. From the remaining 46 studies, 25 were discarded, being descriptive studies without any usable quantitative data. The remaining 21 studies, reported in 23 articles, were regarded for the meta-analysis (6–28). The characteristics of all these studies, one of which is our own, are shown in Table 1. The earliest cohort enlisted patients in 1928 and was published in 1934. Data on mean age at entry and percentage of males were not available for all studies. The two earliest studies had populations selected from mental institutions in the 1920s and comprised mentally ill and retarded patients with epilepsy. This is illustrated in the type of epilepsy, the “epileptic psychosis,” described in these two earlier studies. This selection of patients must be responsible for the very high SMRs, an order of magnitude higher than in all other studies. On these grounds, these studies (6,7) were excluded from further evaluation.
Table 1. Summary of the mortality studies and corresponding determinants
|Author (ref.)||Country||Source population||Type of care||Study design||Case selection||Start of study||Years follow-up||Study size||Age at entry (yr)||Proportionmales||Observed deaths||Expected deaths||SMR|
|Alström (8)||S||I||C + A||FU||I||1924||25||897||23||0.55||160||67.8||2.4|
|Henriksen (11)||S||I||C + A||FU||P||1950||13||2,450||15–89||0.49||94||32.0||2.9|
|Zielinski (15)||EU||C||C + A||FU||P||1965||5||6,710||All||NA||218||121.0||1.8|
|Hauser (17)||USA||GM||C + A||FU||I||1935||29||618||All||0.48||185||81.9||2.3|
|Lühdorf (18)||S||I||A||FU||P + I||1979||6.5||249||72||0.47||121||35.3||3.4|
|Klenerman (19)||EU||I||C||FU||P||1980||11||±300||52 (18–91)||0.65||113||58.3||1.9|
|Lhatoo (20–22)||EU||GM||A||FU||I||1984||8||564||32 (15–59)||0.51||149||58.3||2.6|
|Nashef (23)||EU||I||A||FU||P||1990||3.5||601||33 (10–80)||0.55||24||4.7||5.1|
|Shackleton (24)||EU||I||A||FU||I||1953||41||1,354||19 (1–70)||0.55||403||127.8||3.2|
|Nilsson (25)||S||GMb||A||FU||P + I||1980||13||9,061||54 (15–97)||0.59||4,001||1109.0||3.6|
|Olafsson (26)||S||GM||C + A||FU||I||1960||30||224||NA||0.70||45||28.0||1.6|
|Lindstern (28)||S||GM||A||FU||I||1985||11||107||52 (17–83)||0.57||39||15.7||2.5|
Figure 1 is a graphic representation of the remaining 19 studies and compares the SMRs and the corresponding 95% confidence intervals of patients with epilepsy between the different studies. The SMRs given in the different studies range from 1.3 to 9.3 (8–28). Moreover, the confidence intervals are not overlapping, and the summary interval is disjointed (i.e., not all studies fall within the boundaries of the individual confidence intervals). This indicates that the variation around the pooled estimate cannot be explained by random error alone. Furthermore, the “best possible studies” (i.e., studies that were performed under optimal conditions and using the most favorable methods: from the population or general practitioners using incident cases and with a long follow-up) all give SMRs ranging from 1.6 to 2.6. We formally tested for heterogeneity between the studies. The χ2 test statistic was highly significant (p < 0.001), indicating that the studies estimate different underlying relative risks. As there is heterogeneity between the studies, a summary estimate has no relevant meaning.
Figure 1. Mortality risks and 95% confidence intervals of the 19 studies in the homogeneity test. The best possible studies (numbers 17, 22, 26, and 28: see text) are marked with an asterisk.
Download figure to PowerPoint
A multivariate fixed-effects model was used to explain the source of variation between the different risk estimates. We used the characteristics of the studies as presented in Table 1. The explained variance of the different determinants entered separately into the model, expressed as R2, is given in Table 2, alongside the corresponding p values. Most of these determinants reveal no significant clarification of the variance between the various risk estimates. However, both “source population” and “case selection” explained a significant proportion of the variation in the risk estimates in the model. A graphic representation of the relative risks as a function of these two determinants is given in Figs. 2 and 3. In these figures, the size of the markers reflects the relative weight of the various studies. The determinant “follow-up duration” could explain only <1% of the variation. The relative risk as a function of this determinant is given in Fig. 4. When “source population” and “case selection” were entered simultaneously in a multivariate model, nearly three fourths of the total variance of the estimates could be explained by the combination of these two determinants (R2 = 0.70; p = 0.001).
Table 2. Determinants in the precision-weighted regression model
|Parameter||All studies||Restricted samplea|
|R2||p Value||R2||p Value|
|Type of care||0.11||0.41||0.10||0.46|
|Starting year of study||0.17||0.079||0.02||0.55|
Figure 2. Scatterplot of the mortality risk as a function of the determinant “source population,” entered as categoric variables. The size of the markers represents the relative weight of the corresponding study, large markers indicating studies with high precision. The different kinds of “source population” are illustrated. “Community” studies are those studies derived from the population base and insurance companies; “General medical” studies from both general practitioners and medical clinics; and “Institutional” studies encompass patients originating from epilepsy institutes and neurology clinics.
Download figure to PowerPoint
Figure 3. Scatterplot of mortality risk as a function of “case selection,” given as “prevalent” epilepsy cases; “mixed” incident/prevalent cases; and “incident” cases. The determinants were entered as categoric variables. The sizes of the markers represent the corresponding study's relative weight.
Download figure to PowerPoint
Figure 4. Scatterplot of the mortality risk as a function of “follow-up duration” in years. The size of the markers represents the relative weight of the various studies.
Download figure to PowerPoint
We found less heterogeneity for community studies than for studies among institutionalized epilepsy patients (Fig. 2). Because the studies were still heterogeneous, we decided not to perform a precision-weighted analysis. SMRs in community studies ranged from 1.3 to 3.1, and SMRs for institutionalized populations ranged from 1.9 to 5.1. Moreover, we found more heterogeneity for incident cohort studies, that is, the newly diagnosed patients, than for the prevalent cohort studies (Fig. 3).
As the study of Nilsson et al. (25) is rather large, it is possible that its relatively large weight could have a marked influence on the results. Therefore, after excluding this study, we performed all analyses once more (see Table 2). The variance explained by the determinant “source population” decreased slightly to 40%, whereas “case selection” could now explain only 9% of the variance. Follow-up duration was able to explain 19% of the variance on its own.
- Top of page
All available follow-up studies of patients with epilepsy show an elevated mortality risk. However, the estimates in the studies are highly heterogeneous, indicating that the various estimates show more variation around the underlying mortality risk than can be explained by chance alone. This means that a single mortality risk estimate for patients with epilepsy could not be calculated from simply pooling these studies. In our meta-analysis, the determinant “source population,” which characterizes the setting from which patients were recruited, explained half of the variation in the relative risks between the studies. This also was the case after the exclusion of the large study by Nilsson et al. (25), which dominates the analysis markedly because of its large weight.
In our initial analysis, the determinant “case selection” proved significant and could explain on its own little more than half of the variation. However, after exclusion of the Nilsson study (25), this diminished markedly to <10%. Surprisingly, the distinction between “chronic” patients, or inpatients in long-term institutional care, and “ambulatory,” or outpatients, was in itself nonsignificant in the analysis. Long-term residents are thought of as a highly selected group of patients with severe epilepsy (19), and two studies (11,17) illustrate in subanalyses that the mortality risk increases with increasing severity, recorded as seizure frequency.
The length of follow-up is known to influence relative mortality, being highest in the first 10 years after diagnosis, decreasing as time after diagnosis increases (3,4,17,20,24). In the initial analysis, the “duration of follow-up” could not explain any variation. The graphic representation shown by Fig. 4 illustrates this; the relatively “heavy” studies are evenly distributed over both the X and Y-axes. However, the large study of Nilsson et al. (25), with its relatively short follow-up period, influences the analysis markedly. After this study was excluded from the analysis, the amount of variation explained by the determinant “follow-up duration” increased.
The time period in which the studies were performed ranges from the 1940s to the present, a period that has seen the introduction of many new drug therapies. Nevertheless, the determinant “start of study” could not explain any of the variation in the relative risks. Apparently, both the general population and patients with epilepsy benefit equally from the reduced mortality over time. Study size, as well as country in which the study was performed, also were unable to explain any of the variation. However, it is important to note that all countries are industrialized western countries, with similar medical cultures. There are reports in the literature that mortality is higher in men than in women (3,4,11,15,17,24). This finding, however, is not consistent, and many studies show quite the opposite (14,16,20,23). It is therefore questionable whether this determinant is responsible for any part of the variation of the relative risks. Regrettably, as the sex distribution was not reported for all studies, we were not able to incorporate this determinant in the multivariate fixed-effects models.
“Mean age at entry” into the study is another determinant that is not reported by all studies and is therefore not available for investigation in our analysis. Besides, one is confronted with a conceptual problem as age at entry and age at diagnosis are two of the many different aspects of the determinant age. These two determinants are easily confused; when prevalent cases are involved, they represent different determinants, but when incident cases are considered, they are essentially the same. Reports in the literature illustrate that when age at diagnosis increases, the mortality risk increases (17,20,24). However, many other studies that use determinants such as “age at examination”(11), and “age group”(14,15) in populations of prevalent epilepsy patients generally show a decrease in relative mortality risk as age increases. It therefore remains unclear what part age plays in the variation of the estimates.
Comparisons of SMRs are invalid unless all status-specific study population rates are constant multiples of the specific reference population. We cannot determine whether this assumption holds for our studies, as many studies did not report age and sex characteristics.
Differences in diagnostic criteria that have changed over time and case selection are problems that hamper the comparison of studies. More often than not, authors have neglected to define these essential topics clearly. It seems likely that a large part of the remaining unexplained variability is due to differences in epilepsy diagnosis and classification (29,30). Both have been, and still are, subject to varied implementation and interpretation. The consequences are such that the prognosis of the epilepsies will remain something that we cannot clearly define. We therefore stress the importance of a uniform implementation of the diagnostic instruments used to classify the epilepsies, both for treatment of the disease, and for research purposes.
In summary, we have studied the variation between the different relative risks for epilepsy mortality through determinants that have been signaled as being in part responsible for this variation. Little more than half of the variation between the risk estimates in our analysis had a definable cause. Thirteen of the 19 studies had SMRs between 2 and 4. Given the inherent variability of the various studies, such consistency is quite remarkable. The considerable variation in mortality risk ratios is evident, and more high-quality studies of mortality in epilepsy are needed to signal factors that cause variation of the increased mortality risk. Such studies would pave the way in defining the true representation of the mortality risk summary.