Severity and burden of hand, foot and mouth disease in Asia: a modelling study

Submodel 1: infection rate model

The infection rate model estimates the age-specific proportion of children who must be infected by the EV-A71 virus each year to reach the observed level of seroprevalence. Serological data from China,18–20 Singapore,21 22 Taiwan,23 24 26 Thailand27 and Vietnam28 were extracted from the original papers (figure 2, left), from which it can be observed that seropositivity to EV-A71 starts low at infancy and reaches a plateau around 40%–60% by adolescence, a finding which is corroborated by cord blood seropositivity (indicative of maternal antibody levels) of 55% in Vietnam,28 47%–50% in Taiwan,25 26 38% in Singapore21 and 26% in China.62 To accommodate these patterns, we developed and fit models of the age-specific seroprevalence using seven dynamical models which make varying assumptions for the reason behind these phenomena. These assumptions include (i) an infection rate that drops to zero at age T, beyond which no further infection is possible; (ii) an exponentially decreasing infection rate which is asymptotically zero as age increases and (iii) antibodies for EV-A71 in humans will diminish over time like influenza63 64 and eventually reach an equilibrium from new infections and loss of immunity. The details of each of the models are explained further in online supplementary appendix A1.

Request permission

Posterior distributions of parameters were estimated in R65 using Markov chain Monte Carlo (MCMC) and R package ‘deSolve’66 with non-informative prior distributions. The optimal model to represent infection rate was determined by the fit to the data and the model complexity and was evaluated using the deviance information criteria (DIC).

Submodel 2: symptomatic incidence rate model

Since October 2000, Singapore has adopted a comprehensive approach for HFMD surveillance including a mandatory case reporting system by medical practitioners and childcare centres,67 with children at childcares being screened once or twice daily for symptoms and then subsequently isolated for 10 days if symptomatic. The clearly defined population catchment in the city-state and the dual layers of doctor-driven and teacher-driven surveillance make the data on notified symptomatic presentations of HFMD collected by the Singapore’s Ministry of Health unusually complete compared with the rest of Asia. In calculating symptomatic incidence rate, we make the assumption that all symptomatic paediatric HFMD were clinically apparent and captured by the surveillance system in Singapore.

We constructed a matrix containing the incidence    of cases in year    (2003–2012) among age a (1–18), scaled by the number of individuals in the resident population for that age and year derived from the census.68 To separate the effects of age and time on incidence rate, singular value decomposition was performed on the incidence matrix to find the first left-singular vector    which corresponds to the largest singular value. This vector is normalised to   , which is equivalently the normalised first principal component of   , and represents the systematic variation attributable to age. The second vector is a time-series index,   , representing the stochastic variation attributable to HFMD incidence level each year. Values for years 2013–2015 were estimated using linear regression using data from Singapore’s Weekly Infectious Disease Bulletin.69 The modelled HFMD incidence rate is estimated from the product of the two vectors,   .

The EV-A71-specific incidence is a proportion of all HFMD cases. We obtain   , the year effect due to EV-A71, by taking it as a binomial sample from the total year effect,   . The proportion of cases attributed to EV-A71 in 2008–2015,   , was estimated from data obtained from the National Surveillance Program in Singapore which collects clinical specimens from patients in 107 clinics and 2 hospitals, KK Women’s and Children’s Hospital and National University Hospital (figure 3, right). Throat or vesicle swabs were collected from patients <12 years of age with clinical diagnosis of HFMD. Specimens were stored at room temperature or 4° (preferred) for no longer than 48 hours, before being sent to the National Public Health Laboratory for testing and typing. Data on    prior to 2008 were estimated in a study by Ang et al.52 The average EV-A71 incidence rate by age is the average of the EV-A71 year effect, weighted by the age effect, or   . The model is described in detail in online supplementary appendix B.

Request permission

Submodel 3: hierarchical model of disease severity

The upper levels of the severity pyramid from across Asia were modelled as nested binomial distributions, with the probability of reaching state i or higher given the case reaches the lower state j denoted as   . The states are D (death), S (severe disease, ‘CNS’), H (hospitalised) and C (symptomatic case) so that, for example,    denotes the case hospitalisation rate. Standard rules of conditional probability mean that the product of these probabilities gives the risk of progressing through multiple levels; for instance, the overall case fatality rate (CFR) is   . As a result, the likelihood function for data with missing tiers of the pyramid is specified via combinations of the basic parameters. For example, 4625 hospitalisations and 11 deaths were documented in the Sarawak 1997 outbreak,49 with no mention of severe cases. This data set will allow us to estimate   , and thus make inferences on    and    through the multiplicative relationship, where    is the individual-level estimate for paper p. Detailed description of the model and all data sources can be found in online supplementary appendix C1 (data), C2 and C3 (results) and C4 (model). The posterior distribution of parameters were estimated using the Metropolis Hastings algorithm (MCMC) in R.

Disability-adjusted life-year burden

To estimate the DALY lost across Asia to HFMD, we conservatively estimated the incidence of HFMD infection to be twice the incidence of EV-A71 using the incidence rate model (submodel 2), to allow for infection by other aetiological agents of HFMD which tend to be more common than EV-A71. The number of symptomatic cases was then calculated using the asymptomatic rate from the infection rate model (submodel 1). The numbers of more severe manifestations (hospitalised, CNS and deaths) were then estimated using the parameters from the hierarchical model (submodel 3). Again, posterior distributions were used to characterise uncertainty in these estimates.

Disability weights from GBD201316 for severe (0.133, 0.088–0.19), moderate (0.051, 0.032–0.074) and mild (0.006, 0.002–0.012) acute episodes of infectious disease were used as estimates for HFMD with complication, hospitalisation and symptomatic HFMD, respectively. To estimate the disability duration of each tier, we extracted commonly reported summary statistics such as the sample size, mean, median, range, CIs and variances of the disability duration of their samples from 31 papers and modelled them into Weibull distributions which best describe the statistical properties (online supplementary appendix D). The disability duration for fatalities was determined using the period life expectancy of Japan,70 which represents the expected lifespan of humans as healthcare improves in the current century.

We used the age-weighted approach15 which assigns greater weights to illnesses that occur during periods when the social role is higher, most notably between ages 10 and 55. As disability from HFMD usually occurs among children, this method provides a conservative estimate of DALY.

Under this framework, the formula for DALY for an individual is

where    is the disability weight (1 for premature mortality),   =0.03 is the discount rate,   =0.166 is the age-weighting correcting coefficient,   =0.04 is the parameter from the age-weighting function,    is the age of onset and    is the duration of disability or time lost due to premature mortality.

Submodel 1: serological evidence for EV-A71 infection rate

Among the seven contesting models, model 7 with an unconstrained initial rate of infection and a decreasing rate of infection that reaches 0 at time T has the best fit to the data. The model formulation is

where    represents the seropositive rate at time t, c and k are positive real numbers and    is the rate which seropositivity increases (due to new infections).

The fitted values are c = 2.3 (95% CI 1.47 to 5.28), k = 4.79% (95% CI 1·92% to 8.07%), T = 9 years (95% CI 9 to 10), which translates to around 10% required infection rate in the first year of life, decreasing to 5% at ages 5–6, and drops to infinitesimally near 0% at around age 9 (figure 2, mid). Details of all models, their estimated parameters and the DIC can be found in online supplementary appendix A1, A2 and A3.

Submodel 2: incidence analysis within Singapore

The year effect,   , which captures the yearly variation of total HFMD incidence (figure 3, mid) has a correlation >99% with the overall number of cases of HFMD in Singapore. There is a positive slope of gradient 0.026 (95% CI 0.0003 to 0.052) indicating an increase in number of cases over time. The year effect of EV-A71 is estimated using laboratory data (figure 3, right) and does not have a significant trend with slope −0.0091 (95% CI −0.033 to 0.015). Year effects are presented in online supplementary appendix table A4.

The age effect    (online supplementary appendix table A5) which captures the systematic variation of incidence attributable to age is estimated from data from years 2005 to 2012 and is representative of the proportion of total incidence for each age, showing that >50% of disease incidence are due to children <4 years of age. Around 2.2%–2.9% of the population is affected by EV-A71 each year during their younger ages (1–3), but this quickly reduces to 1.2% by age 5 (figure 2, mid).

For Singapore, the required rate for new infections is much higher than the actual observed level, and the differences are due to asymptomatic infection. For the most vulnerable age group, around 70% of EV-A71 cases do not show sufficient symptoms to be identified either at presentation to their preschool or by their doctor (figure 2, right). Given the high intensity of Singapore’s notification system, this is arguably a lower bound on how much unnotified infection there is across other Asian countries.

Submodel 3: hierarchical model of disease severity

For HFMD in general (ie, without reference to the specific viral agent), the symptomatic case hospitalisation ratio is estimated to be 6% (2.8%–14.9%). Among hospitalised cases, 18.7% (6.7%–31.5%) develop CNS complications, of which 5.0% (2.9%–7.4%) were fatal. Overall, the CFR for HFMD is estimated to be 52.3 (24.4–92.7) per 100 000 symptomatic infections. Among the papers which identified EV-A71 as the pathogen, the majority recruited children seeking treatment in hospitals. Of these, 36.9% (25.9%–48%) were associated with CNS complications, of which 10.5% (4.9%–17.8%) died. In contrast, there were no discernible differences in the fatality rate of HFMD patients with CNS complications between cases with EV-A71 and all HFMD infections, possibly because most severe cases were indeed caused by EV-A71. Estimates within studies and overall are presented in figure 4 and table 1, which show a summary of all estimated severity quantities of HFMD in Asia with their associated uncertainties.

Request permission

Disease burden

The disease burden estimated using our approach is 96 900 (40 600–259 000) DALYs per annum in eight high-endemic countries in East and Southeast Asia (estimates by country are provided in table 2). Our approach is conservative and will represent the minimum expected burden of disease. Despite the low CFR, the majority of disability comes from mortality (~99%) because deaths from those aged 1 to 3 years contribute very long disability lengths. In contrast, the disability length for mild symptomatic HFMD is estimated at 6.0 (5.8–6.2) days; hospitalisation about 8.3 (8.1–8.5) days, while HFMD with complications at (10.4–11.1) days.