Using healthcare-seeking behaviour to estimate the number of Nipah outbreaks missed by hospital-based surveillance in Bangladesh

Abstract

Background

Understanding the true burden of emergent diseases is critical for assessing public-health impact. However, surveillance often relies on hospital systems that only capture a minority of cases. We use the example of Nipah-virus infection in Bangladesh, which has a high case-fatality ratio and frequent person-to-person transmission, to demonstrate how healthcare-seeking data can estimate true burden.

Methods

We fit logistic-regression models to data from a population-based, healthcare-seeking study of encephalitis cases to characterize the impact of distance and mortality on attending one of three surveillance hospital sites. The resulting estimates of detection probabilities, as a function of distance and outcome, are applied to all observed Nipah outbreaks between 2007 and 2014 to estimate the true burden.

Results

The probability of attending a surveillance hospital fell from 82% for people with fatal encephalitis living 10 km away from a surveillance hospital to 54% at 50 km away. The odds of attending a surveillance hospital are 3.2 (95% confidence interval: 1.6, 6.6) times greater for patients who eventually died (i.e. who were more severely ill) compared with those who survived. Using these probabilities, we estimated that 119 Nipah outbreaks (95% confidence interval: 103, 140)—an average of 15 outbreaks per Nipah season—occurred during 2007–14; 62 (52%) were detected.

Conclusions

Our findings suggest hospital-based surveillance missed nearly half of all Nipah outbreaks. This analytical method allowed us to estimate the underlying burden of disease, which is important for emerging diseases where healthcare access may be limited.

Key Messages

• Surveillance of emerging infections often relies on hospital systems that only capture a minority of cases, though knowing the true burden of these diseases is critical for assessing public-health impact.

• Using Nipah virus in Bangladesh as an example, we demonstrate how healthcare-seeking data can estimate true burden.

• We found only 52% of Nipah outbreaks were detected by surveillance.

Introduction

Emerging infectious diseases are an important public-health threat^1–4 and, for those with the potential for sustained person-to-person transmission, the risk is global.^1,5 For such infections, knowing the true incidence of cases is important for estimating the burden of disease, tracking changes over time and informing public-health surveillance programmes designed to provide early warnings of outbreaks. Surveillance systems often use hospital-based surveillance for case detection,^6,7 which means that disease burden is commonly underestimated because not all patients with the disease seek care at a hospital, if at all.⁸ Reasons for reduced healthcare-seeking include poverty, lower education,^9,10 milder disease, being a woman^11,12 and increased distance to a healthcare facility.^13–15 In low-resource settings, particularly, poor transportation infrastructure contributes to limited access to healthcare.

Although public-health practitioners often assume cases are missing from surveillance data, few tools exist to estimate the number of cases missed. In 1988, the US Centers for Disease Control and Prevention (CDC) published Guidelines for Evaluating Public Health Surveillance Systems (updated in 2001), which aimed to efficiently and effectively standardize evaluations of public-health surveillance systems using a series of broad characteristics, including sensitivity and representativeness.¹⁶ Sensitivity is defined as the proportion of true cases or outbreaks detected by the surveillance system and representativeness as the accurate description of cases over time and their distribution in a population by place and person. However, to measure both sensitivity and representativeness, and validate the data collected by the surveillance system, external data are required to compare and determine the true incidence of disease in the population. Such data can include medical records and registries, which seldom exist or are incomplete in low-resource settings, highlighting the need for other methods to evaluate these surveillance characteristics.

A part of the family Paramyxoviridae, Nipah virus causes severe encephalitis in human infections, with the case-fatality ratio ranging between 40 and 70%.¹⁷ In Bangladesh, Nipah outbreaks have been identified almost yearly since 2001 in rural communities in the central and northwest regions.^17,18 Human infection occurs either through transmission from Pteropus bats, the natural reservoirs or from subsequent human-to-human transmission.^17,19 Human consumption of date palm sap contaminated by infected bat urine or saliva has been identified in Bangladesh as the primary route of transmission from bats to people.¹⁹ Similar to case identification of other rare emerging infectious diseases, Nipah-case identification is important because it offers intervention opportunities to prevent human-to-human transmission and large-outbreak propagation, and to track genetic changes in, and transmissibility of, the virus.

Nipah-case detection in Bangladesh relies largely on a hospital-based surveillance system, established in 2007, made up of a network of three public tertiary-care hospitals.²⁰ The vast majority of recorded outbreaks since 2007 have been detected through this system, with the remainder identified through media surveillance or ad hoc detection at other hospitals, whereby physicians at other hospitals that are not a part of the surveillance system sometimes report and refer suspect cases.²¹ Despite the substantial effort exerted in identifying Nipah cases in surveillance hospitals, cases are likely missed in the current detection system. Many rural families may not be able to easily access hospital care^14,22 and 49% of patients in Bangladesh seek care from the private sector.²³ Although resources to detect cases are scarce, if most outbreaks are missed by this surveillance system, additional investments to find Nipah cases could help ensure we meet the public-health objectives for understanding the epidemiology and preventing large outbreaks. Here, we used detailed data of healthcare-seeking behaviour of individuals with severe neurological disease to estimate the total number of Nipah outbreaks that occurred in Bangladesh from 2007 to 2014, allowing us to estimate the number of outbreaks missed by the hospital-based surveillance system.

Methods

We assumed that whether a case was captured by the surveillance system depended on the individual’s healthcare-seeking behaviour, which differs by where they live in relation to a tertiary-care surveillance hospital and the severity of their disease. Cases may be missed if individuals live far from the surveillance hospitals (resulting in attending other hospitals where they would not be captured as Nipah cases or not attending hospital), if they are severely ill because they are too sick to travel or if they lack means of transportation, have poor road infrastructure or limited funds to get to a hospital. We used logistic-regression models to characterize the probability of presenting at a surveillance hospital as a function of both disease outcome (i.e. whether the individual died or not) and distance from a surveillance hospital for individuals with Nipah-like disease (i.e. those suffering from neurological disease) using data from a healthcare-seeking study in Bangladesh. We used the results of the healthcare-seeking model to estimate the number of missed Nipah outbreaks in the country using the observed spatial distribution of Nipah outbreaks.

Nipah-case surveillance

In 2007, the Institute of Epidemiology, Disease Control and Research, the International Centre for Diarrhoeal Disease Research, Bangladesh (ICDDR, B) and US CDC established Nipah sentinel sites at three 500-bed tertiary-care hospitals in the districts of Rajshahi, Rangpur and Faridpur to systematically test encephalitis patients for evidence of Nipah virus. In each of these three sentinel surveillance hospitals, patients hospitalized with fever plus altered mental status or seizures from December to March, the known Nipah season,¹⁸ had blood and sometimes also cerebrospinal fluid sampled and tested for anti-Nipah IgM by ELISA.²⁰ Patients with IgM antibodies were considered laboratory-confirmed. Any laboratory-confirmed case resulted in an outbreak investigation to identify additional cases in the community. The home location of laboratory-confirmed cases was recorded through hand-held GPS devices. The index case for each outbreak was the first patient who presented at the hospital. The Bangladesh Ministry of Health and Family Welfare authorized protocols. All interviewed individuals provided informed consent; for those <18 years of age, individual and parental consent was obtained.

Probability of neurologic-disease case detection

To estimate the probability that a person with severe encephalitis would present at a surveillance hospital, we used the results of a large healthcare-utilization study from three districts in Bangladesh (Rajshahi, Khulna and Chittagong).^7,24 This study used data from household surveys conducted in the areas surrounding the tertiary-care surveillance hospital in each of the three districts. The methods for estimating the probability of detecting a case of severe neurological disease are published elsewhere.⁷ In brief, community-based healthcare-utilization surveys were used to identify patients with severe, acute neurological illness, compatible with Nipah-virus encephalitis, and capture information about whether each patient sought care at the nearest surveillance hospital and whether the person died. We fit a logistic-regression model to estimate the probability that a patient presented at the tertiary-care hospital. As healthcare-seeking may differ by disease severity, we included an interaction term for whether or not the patient died:

$$\mathit{logit}\left(\mathit{Pr}(Y_i=1|d_i{,x}_i)\right)={\beta }_0+{\beta }_{1\mathrm{ }}{d}_i+{\beta }_{2\mathrm{ }}{x}_i+{\beta }_{3\mathrm{ }}{d}_i{x}_i,$$ [1]

where Y is whether the individual attended a tertiary-care hospital, d is the Euclidean distance between the case and the closest tertiary-care hospital, and x is whether the individual died (coded as 1 when they died and 0 if they survived).

Estimation of number of missed Nipah outbreaks

We assumed that the total number of Nipah outbreaks within Bangladesh is equal to the sum of the total number of index cases at each distance from a surveillance hospital. We also separately considered those index cases who died and survived:

$$N_{\mathit{tot}}=\sum _{d,x}{(n}_{\mathit{tot}}\left(d,x=0\right)+n_{\mathit{tot}}\left(d,x=1\right)),$$ [2]

where $n_{\mathit{tot}}\left(d,x\right)$ is the total number of Nipah index cases at distance d from a surveillance hospital with outcome x, which will be made up of the sum of observed cases and unobserved cases:

$$n_{\mathit{tot}}\left(d,x\right)=n_{\mathit{obs}}\left(d,x\right)+n_{\mathit{unobs}}\left(d,x\right).$$

If we assume that the probability of observation is dependent on the distance to the hospital, we can rewrite this as:

$${E(n}_{\mathit{tot}}\left(\mathrm{d},x)\right)=n_{\mathit{obs}}\left(d,x\right)+\left(1-P_{\mathit{obs}}\left(d,x\right)\right){\cdot n}_{\mathit{tot}}\left(d,x\right),$$

where $P_{\mathit{obs}}$ represents the probability of a Nipah index case being observed (i.e. patient attending the surveillance hospital) as distance d and outcome x and can be estimated using Equation 1. Therefore, the total number of index cases at distance d with outcome x is:

$$n_{\mathit{tot}}\left(d,x\right)=\frac{n_{\mathit{obs}}\left(d,x\right)}{P_{\mathit{obs}}(d,x)}.$$

To estimate the total number of Nipah outbreaks, we first calculated the Euclidean distance to the nearest surveillance hospital for the index case of each Nipah outbreak detected through the hospital-based surveillance system between 2007 and 2014, and identified whether the patient detected at the hospital died. We estimated $n_{\mathit{obs}}\left(d,x\right)$, the probability of observing a Nipah outbreak at each kilometer, d, between 0 and 120 km from a surveillance hospital, by initially fitting a smooth spline to the empirical cumulative distribution function of the number of index cases at each distance (Supplementary Figure 1, available as Supplementary data at IJE online). We did this separately for Nipah patients who survived and those who died. We then used Equation 2 to calculate the total number of Nipah index cases or outbreaks throughout this period. We used 120 km as the maximum distance, as only communities within this distance were included in the healthcare-utilization study. There was only one detected Nipah case (detected through the hospital surveillance system or otherwise) beyond this distance during the study period, suggesting the risk of Nipah was minimal at greater distances. We used bootstrapping to quantify uncertainty, where both the Nipah index cases and the cases in the hospital utilization study were randomly re-sampled with replacement of over 2000 iterations with a logistic-regression model refit in each iteration. This approach therefore incorporated sampling uncertainty both in the Nipah cases and in the model-fitting process. We calculated 95% confidence intervals (CIs) from the 2.5 and 97.5% quantiles of the resultant distribution. We conducted a logistic regression to test an association between distance to a surveillance hospital and size of the detected outbreak to rule out inclusion of this parameter.

We additionally performed three sensitivity analyses, detailed in Supplementary Text 1, available as Supplementary data at IJE online. We assessed the sensitivity of our results to using healthcare-utilization data from only within the area where Nipah outbreaks occur, considered each individual case for the calculation of $n_{\mathit{obs}}\left(d,x\right)$ instead of spillover events and conducted a simulation study to demonstrate that this approach can estimate the true number of outbreaks, even in situations of spatial heterogeneity in risk. All analyses were conducted in R (v. 3.4.2, R Foundation for Statistical Computing, Vienna, Austria).

Results

From 2007 to 2014, there were 62 detected outbreaks of Nipah virus in Bangladesh, consisting of 145 cases. Most outbreaks (72%) consisted of a single case. There was no association between distance to a surveillance hospital and the size of the detected outbreak (P = 0.37). The mean cluster size per index case was 2.3 (median cluster size of 1 case per index case). Index cases were clustered in the central and northwest regions of Bangladesh, with those who did not visit a surveillance hospital scattered in the same general region, except one on the eastern side of Bangladesh (in Comilla district) (Figure 1A). Fifty-five outbreaks (89% of all outbreaks) and 110 cases (76% of all cases) were detected by the surveillance hospitals, with the remainder detected through other means, such as media reports. There were substantial differences in the number of outbreaks detected by distance from hospital, with a majority of detected outbreaks located within 30 km of a surveillance hospital (Figure 1B). Eighty-nine per cent (49/55) of the patients identified as index cases who were detected at a hospital died, compared with 76% of non-index cases (68/90), (p = 0.054).

Figure 1.

(A) Map of 2007–14 index cases of patients with Nipah virus in Bangladesh who did and did not visit a surveillance hospital and location of Nipah surveillance hospitals. (B) Histogram of number of Nipah outbreaks by distance to a surveillance hospital by those who survived and died. (C) rho(d) prediction and model fit with 95% confidence intervals. (D) Cumulative number of observed and estimated unobserved Nipah outbreaks in Bangladesh 2007–14.

There were 180 cases of severe neurological disease identified in the healthcare-utilization survey. Forty-four of these individuals with severe neurological disease died, 61 visited a surveillance hospital from the healthcare-utilization study for care and 85 were from the district of Rajshahi, where Nipah outbreaks have been identified. Among the patients who did not visit a surveillance hospital, the majority sought care within the informal sector (e.g. traditional healers, homeopaths and pharmacies) or at hospitals outside the surveillance network.⁷

The logistic-regression model (Table 1) and probability distribution describing healthcare-seeking behaviours of these individuals (Figure 1C) demonstrated that patients with neurological disease who eventually died were more likely to have attended a surveillance hospital than those who survived (55 vs 27%). For Nipah index cases, the odds of attending a surveillance hospital were 3.2 (95% CI: 1.6, 6.6) times greater for patients who eventually died (i.e. who were more severely ill) compared with those who survived. However, the probability of attending a surveillance hospital also changed as a function of distance from hospital, falling from 82% (95% CI: 58, 94) at 10 km from the hospital to 54% (95% CI: 38, 69) at 50 km for those who subsequently died. For patients who survived, the probability of seeking care fell from 36% (95% CI: 24, 51) at 10 km to 25% (95% CI: 18, 34) at 50 km. When controlling for distance, the adjusted odds of attending a surveillance hospital were 10.0 (95% CI: 2, 60) times greater for patients who eventually died compared with those who survived.

Table 1.

Logistic model results of the probability of attending a surveillance hospital by distance and outcome, 2007–14

	OR (95% CI)
Intercept	0.67 (0.30, 1.35)
Death as outcome	10.0 (2.01, 60.3)
Distance	0.99 (0.97, 1.00)
Distance $\times \mathrm{ }\mathrm{d}$eath	0.98 (0.95, 1.01)

Applying our characterization of healthcare-seeking behaviours to the observed distribution of Nipah cases, we estimated there were 119 Nipah outbreaks between 2007 and 2014 (95% CI: 103, 140) (Figure 1D), at an average of 15 outbreaks per Nipah season. Sixty-two of these outbreaks were detected (Supplementary Figure 2, available as Supplementary data at IJE online). This represents a detection probability of 46% from the hospital-based surveillance system (95% CI: 39, 53) and 52% (95% CI: 44, 60) by any detection method. If we assume a mean number of Nipah cases of 2.3 per outbreak (consistent with the mean number of cases in each observed outbreak over the study period), this means that, between 2007 and 2014, there were 274 cases, at an average of 34 per season. The sensitivity analysis that restricted our analysis to using probability-of-detection estimates from Rajshahi district only gave consistent results (119 outbreaks, 95% CI: 100, 146) (Supplementary Figure 3, available as Supplementary data at IJE online). The sensitivity analysis that used all Nipah cases in our analysis also yielded similar results for the total number of Nipah cases estimated between 2007 and 2014 (279 cases, 95% CI: 248, 315) (Supplementary Figure 4, available as Supplementary data at IJE online). Finally, using simulations, we demonstrated that our approach recovers the true number of outbreaks, even where there exists spatial heterogeneity in risk and spatial heterogeneity in the observation process (984, 95% CI: 914, 1061; compared with a true number of 1000) (Figure 2).

Figure 2.

Results of simulations. (A) Simulated locations of outbreaks in a 100 × 100-km area with spatial heterogeneity in risk (as generated through a Matern cluster process). (B) Probability of an outbreak being detected depends on the distance at which the outbreak occurred from the hospital [triangle in (A) and (C)]. In addition, no outbreaks were detected in the topleft quadrant [striped shaded area in (A) and (C) to represent outbreaks in a foreign country]. (C) Observed outbreaks. (D) Histogram of the estimated number of outbreaks (excluding those in the striped shaded region) over 100 simulations. The true number of outbreaks was 1000 each time.

Discussion

Our analysis provides an example of how healthcare-utilization data can be used to evaluate surveillance sensitivity and representativeness—critical components of the widely used CDC evaluation guidelines. We estimated that 119 Nipah outbreaks and 274 total cases occurred in Bangladesh from 2007 to 2014—more than double the number of outbreaks detected by the hospital-based surveillance system. This represents an average annual burden of 34 cases from 15 outbreaks. Media-based and ad hoc surveillance efforts captured seven outbreaks not detected through the hospital-based surveillance, and therefore these systems can only make up a small part of this shortfall. Our results indicate the Nipah surveillance system is not sensitive and not representative of the underlying spatial distribution of Nipah cases. This method of using healthcare-seeking data to estimate the number of missed cases is a useful strategy for evaluating hospital-based surveillance systems when no other external data exist. If about half of all Nipah cases in Bangladesh are missed by surveillance, the ability of the public-health system to quickly identify and respond to Nipah outbreaks remains limited and additional strategies to improve case detection should be explored.

We found that individuals with severe neurological disease were markedly less likely to visit surveillance hospitals at distances greater than 50 km. In 2005, the WHO revised the International Health Regulations (IHR), a major advancement in global disease surveillance,²⁵ whereby member states are expected to strengthen capacity for the detection, reporting and control of international public-health risks.²⁶ Despite substantial effort and resources to build capacity for Nipah-virus outbreak detection in Bangladesh, gaps remain due to inadequate access to healthcare facilities. Suboptimal access to hospitals is not unique to Bangladeshi populations; indeed, it is a commonly identified barrier to improved health in many low- and middle-income countries.^9,10,12–15 These countries might face difficulties in meeting IHR requirements if their surveillance systems continue to rely primarily on patients seeking care at hospitals.

Our analysis used distance to hospital to estimate the number of missed cases because this has been repeatedly shown to be associated with healthcare-seeking.^7,27,28 As there was no association between the size of outbreaks we detected and distance from a hospital, the size of an outbreak was not accounted for in our initial estimation. However, we obtained similar results when all Nipah cases were considered, rather than just the index cases. Ultimately, the cases missed by hospital-based surveillance reflect limited access to tertiary-level healthcare in this population. Whereas efforts to increase access to appropriate care are important, improving surveillance may include working to engage other healthcare facilities in the interim. Exposure-based surveillance, which focuses laboratory testing only on patients reporting exposure to date palm sap or exposure to other persons with encephalitis, could be used to expand surveillance with marginal costs. This method was sensitive and specific during the Nipah season.⁶

This study used information obtained from a detailed healthcare-seeking-behaviour study to be able to quantify the underlying number of outbreaks. We would not be able to apply this method in settings where there is no information on these behaviours. This highlights the importance of obtaining a quantitative understanding of how people behave when they seek care and how this changes by disease-specific (e.g. disease severity) and individual (e.g. distance from hospital) characteristics. Such behaviour studies are also key to assessing the performance of surveillance programmes.⁷

We estimate that about half of outbreaks were missed. However, the individual risk of contracting Nipah virus remains low. The total population in the 22 districts that have ever reported a Nipah case is 44.8 million, suggesting the rate of becoming infected is <5 per 100 000 persons per year. Although the current burden of disease for Nipah virus is low, the high case-fatality ratio and potential for human-to-human transmission highlight the need to monitor transmission patterns from bats to people and from person to person to identify changes in the epidemiology of disease.⁵ Animal models have suggested that the Bangladeshi Nipah strain is more pathogenic than its Malaysian counterpart. Each additional spillover from bats into humans presents an opportunity for increases in human-to-human transmissibility.²⁹ When recognition of outbreaks is delayed, the public-health response risks being too slow to limit person-to-person transmission. Additional investments in case detection could reduce the risk for wider regional or global spread. With an increased ability to identify Nipah cases, we could further the development of prevention tools, like vaccines, as historic case-detection rates for Nipah virus in Bangladesh are too low to confidently power efficacy trials. If surveillance could identify the estimated half of cases that are currently missed, these studies may be possible.

A limitation of this study is that it assumed patients with Nipah virus have the same healthcare-seeking patterns as patients with severe neurological illness identified in the healthcare-utilization surveys we used to estimate the missing outbreaks.²⁴ Most of the individuals reporting illness in these surveys likely had non-Nipah infections or may have had non-infectious cases of neurological illness; however, the signs and symptoms of encephalitis are the most important drivers of care-seeking and are similar, regardless of aetiology. In addition, though it is conceivable that patients with Nipah virus have different healthcare-seeking behaviours than patients with other severe neurological illness, because they tend to occur in clusters, our estimates are based on Nipah index cases, which occurred as isolated events. We also only considered distances up to 120 km around each surveillance hospital for our analysis. Though this was somewhat arbitrary, only one outbreak (in Comilla district) has ever been reported outside this zone and one outbreak detected between 80 and 120 km from a surveillance hospital. If a considerable number of people infected with Nipah have gone undetected outside the areas included in our analysis, then we have systematically underestimated the number of Nipah cases missed by surveillance in Bangladesh. As our estimate of missing outbreaks is based on observed outbreaks in Bangladesh, we are only estimating missed outbreaks within Bangladesh though the 120-km radii around two surveillance hospitals that cross into India. Importantly, our simulation demonstrates that our approach is robust to spatial heterogeneity in risk and spatial heterogeneity in the observation process.

Although there have been substantial improvements in international surveillance programmes and evaluation, rigorously evaluating the performance of these systems would support the objectives of the IHR.^1,30 We recommend that cross-sectional healthcare-utilization surveys be included in the CDC surveillance evaluation guidelines as additional tools for determining sensitivity and representativeness. Countries that are dedicated to understanding the limitations of their surveillance could follow the example of Bangladesh and invest in healthcare-utilization surveys to gain better insights into surveillance performance.