Our analyses provide estimates of the incidence of encephalitis in England attributable to infectious and noninfectious causes. We present a unique application of capture–recapture models to estimate the number of cases of encephalitis by using an original dataset of well-defined and extensively investigated cases of encephalitis.

Multiple scenarios were investigated to assess the sensitivity of the estimates to various assumptions. Depending on the scenario, estimated incidence ranged from 2.73 cases/100,000/year to 8.66 cases/10⁵/year; all estimates were higher than the 1.5 cases/100,000/year previously reported (3). This unique study has brought together 2 distinct datasets to help address the inevitable limitations within such data sources, particularly those encountered with complex syndromes such as encephalitis. We consider our capture–recapture estimate of 5.23/100,000/year (assuming 54% PPV for HES data) to be the best estimate of encephalitis incidence in England; this is 3.5 times higher than that previously described by Davison et al (3). Our incidence analyses update the Davison et al. estimates; diagnostic advances, emerging etiologic agents, and introduction of new interventions and control strategies (e.g., vaccines) are all likely to have affected incidence estimates over time. Furthermore, our data included both infectious (not just viral) and noninfectious causes of encephalitis, in line with the increased recognition of new immune-mediated encephalitis etiologies.

A higher incidence (adjusted for year, age, and region) of encephalitis was observed among male patients, which is consistent with previous studies (12–16). We also observed higher incidence of encephalitis among patients <1 and >65 years of age. Hyporesponsiveness of the immune system in early life and later immunosenescence render these groups more susceptible to infection, to reactivation of latent infection, or development of encephalitis once infected (17).

Multiple admissions were included in our analyses only if they represented different etiologies or occurred >6 months apart, either of which is not a common occurrence in encephalitis as supported by our data. When multiple occurrences were excluded, incidence was unchanged. Infectious episodes of encephalitis are unlikely to recur in the absence of immunosuppression; relapses in immune-mediated cases are more frequent and were documented in the PHE study.

We extended our encephalitis incidence analyses beyond the use of HES and linked HES data to PHE study data and applied capture–recapture models. The number of incident encephalitis admissions recorded in HES was considerably higher than the number of cases included in the PHE study, even after restricting HES diagnoses to the primary diagnostic field. Conversely, nearly half of the PHE cases were not captured in HES. The poor agreement between these 2 data sources could have several possible reasons.

The likely explanation for nonmatched PHE cases is that testing in the PHE study went far beyond routine clinical practice, which highlighted the extent to which encephalitis can be underdiagnosed. The higher proportion of bacterial cases in the PHE study, classified as meningoencephalitis, suggests that these cases may be coded as meningitis rather than encephalitis in HES. Also, patients with unusual signs and symptoms, such as those with N-methyl-D-aspartate receptor-antibody encephalitis, which typically causes psychiatric symptoms, might not have been classified as encephalitis case-patients in HES (18). Unfortunately, we could not identify the HES codes used for the unmatched PHE cases because the HES data for this study included only patients with an encephalitis code.

Unmatched HES cases likely include true encephalitis cases not reported to the PHE study team and nonencephalitis cases misdiagnosed as encephalitis. Underascertainment in the PHE study is likely, as employing >1 research nurse per region to actively identify cases was not financially feasible: some centers relied on case notification by hospital staff alone. The likelihood that HES admissions coded as encephalitis included misdiagnoses of syndromes with signs that mimic encephalitis is highlighted by the PHE study, in which only 54% of suspected encephalitis patients initially screened during the 2-year period were ultimately included following a rigorous diagnostic process (10). The higher proportion of cases of unknown etiology in unmatched compared with matched HES admissions and their shorter length of hospital stay supports the possibility of misdiagnosis and suggests a more likely diagnosis of a mimicker syndrome such as septic encephalopathy.

Other reasons for mismatches need consideration. The catchment areas covered by the 2 data sources differed slightly, but the results of the capture–recapture model did not change when we excluded the 2 trusts that had hospitals with potential extra encephalitis cases in the HES data. The higher number of admissions in HES could be due in part to the inclusion of patients with postencephalitic sequelae, who would not have been notified to the PHE study. This finding is supported by the higher proportion of cases admitted electively and treated under neurosurgery in unmatched HES admissions; alternatively, some of these cases could be miscoded nonencephalitis mimics, i.e., cases of nosocomial meningitis following surgery.

Three assumptions required for valid capture–recapture estimates also need consideration. First, for a given source every patient should have an equal chance of being ascertained by that source, although different sources may have different probabilities of identifying an individual case. Our analyses stratified by age or region of residence did not indicate any bias in the point estimate linked to these variables. Other variables may have influenced the probability of a patient appearing on a list, such as ethnicity, sociodemographic factors, or heterogeneity in coding between hospitals within regions. Because of the small sample sizes, we could not stratify by these variables. We did find evidence of heterogeneity in etiology, which suggested that in 1 or both datasets, patients with bacterial encephalitis had a different probability of being identified than did patients with encephalitis of other etiologies, which may have led to an overestimate of the number of encephalitis cases. Again, performing a stratified analysis is not easy, both because of small numbers and because we know misclassification and missing data about etiology were present and no pathognomonic features exist to allow different causes to be distinguished in these cases.

Second, the registers used in capture–recapture should be independent; having an encephalitis-specific diagnosis in HES should not affect being included in the PHE study or vice versa. A patient with suspected encephalitis seen by a hospital clinician was likely to be coded as such in HES and also be notified to the PHE study. Thus, these sources are likely to be positively dependent, and the capture–recapture will have underestimated the true number of cases. A less likely scenario is that the datasets were negatively dependent, for example, if encephalitis cases with bacterial etiology were simultaneously more likely to be excluded from the HES data and more likely to be included in the PHE data. This situation would have led to an overestimate in the number of cases. We did not have access to a third data source to evaluate the independence of data sources (19). Nevertheless, 2-source capture–recapture can indicate an upper or (as is likely here) the lower bound of estimates when the direction of dependency is known or highly suspected (20).

Third, no false positive cases should occur due to misdiagnoses or miscoding. As discussed above, coded HES encephalitis cases are likely to have included mimicker syndromes, which would inflate the capture–recapture estimate. We addressed this possibility by applying a range of PPVs to the HES data; even after assuming a PPV as low as 30% for HES-only admissions, the number of cases was still higher than previous estimates. A review of medical records would be necessary to determine the degree of misdiagnosis of true encephalitis cases in HES data. We could not do this because HES do not keep patients’ names, and use of HES data are subject to strict protocols to prevent identification of individual patients.

With a mean length of hospital stay of 34 days, an incidence of 5.23 cases/100,000/year (“best estimate”) equates to 90,852 bed-days of hospital occupancy. On the basis of a bed-day cost of £261 (US$394 million), the cost to the National Health Service would be >£23 million (US$35 million) per year (21). An incidence of 8.66 cases/100,000/year, our maximum estimate, would cost almost £40 million (US$60 million) per year. The actual cost is likely to be higher as patients often require intensive care, costly investigations, and in-patient rehabilitation. Additional costs include long-term care and loss of productivity among many working-age survivors.

In summary, the different scenarios used in this study provide strong evidence that the incidence of encephalitis is higher than that previously estimated in England. This higher incidence has clinical, research, and public health implications. A diagnosis of encephalitis should be considered for patients with compatible symptoms, especially given the increased recognition of immune-mediated encephalitides for which treatment is available and effective if instigated early. Early recognition is important to help reduce the substantial economic and societal costs of encephalitis suggested by our study. Stand-alone HES data are used extensively for public health research; our analyses highlight the extent to which HES-only data might over- or underascertain cases of complex syndromes and the advantages of linking these data to other sources to improve incidence estimates. Encephalitis incidence in this study was higher than that of other neurologic conditions, such as meningococcal meningitis and motor neuron disease, both of which have a higher profile and public focus (22–24). This study highlights the importance of accurate diagnosis and coding for complex syndromes with multiple etiologies to obtain accurate estimates of incidence and to further explore the epidemiology and outcomes of this devastating neurologic illness.