Retrospective analysis assessing the spatial and temporal distribution of paediatric acute respiratory tract infections in Ho Chi Minh City, Vietnam

Data sources

Data for this study were derived from the electronic hospital databases of all paediatric (<16 years) inpatient admissions to three large tertiary hospitals located in central HCMC: Children’s Hospital 1 (CH1), Children’s Hospital 2 (CH2) and the Hospital for Tropical Diseases (HTD). Data from the two Children’s Hospitals were collected between 2005 and 2010 (inclusive) while the data from HTD were available from 2008 to 2010 only.

ARIs are caused by a range of differing pathogens and may have different spatiotemporal patterns and associations with climate and other covariates. Therefore, we allocated respiratory infections into two major clinically differentiable groups based on the International Classification of Disease version 10 (ICD-10) codes version 2016 outlined by the WHO: (1) URIs were amalgamated under the ICD-10 codes J00:J06 and J9:J11 for influenza-like illnesses (ILI) and other acute URIs and (2) LRIs which incorporated the ICD-10 codes J12:J18 and J20:J22 for pneumonia and other acute LRIs. For the purposes of this study, the data from patients with the selected ICD-10 codes who resided in HCMC were extracted and analysed. Each patient record included data regarding age, sex, date of admission and discharge, ICD-10 code at discharge and residential address.

HCMC is a large economic centre in the South of Vietnam with a population of >8 million people. Geographically, it is divided into 24 districts (land areas ranging from 4.18 km²to 704.22 km²) which are subdivided in 322 smaller administrative wards. There are five rural districts (population density ranging from 100 persons/km² to 3326 persons/km²) and 19 urban districts (population density ranging from 2360 persons/km² to 45 582 persons/km²). There are at least one local (district) hospital in each district, which serves both outpatients and inpatients of all ages. However, most local hospitals, especially those in urban districts receive mostly adult outpatients. There are two central hospitals (CH1 and CH2) that specialise in serving the paediatric population in HCMC. Another further central hospital (HTD) specialises in infectious diseases and also receives paediatric patients. While paediatric outpatients in HCMC may choose to visit local or private healthcare facilities or tertiary hospitals, most paediatric patients requiring hospitalisation in HCMC are referred or self-referred to one of the three study hospitals. Therefore, the three study hospitals (CH1, CH2 and HTD) receive the majority of paediatric inpatients from HCMC.

There are two main seasons in HCMC: a dry season from December to April and a rainy season from May to November. Population and district characteristic data (urban vs rural districts, district population density, district economic indicators (expenditure of district budget, revenues of district budget), district agricultural indicators (number of pigs in herds >2 months of age by district, percentage area of district used as rice paddy) and district education indicator (number of primary and secondary schools per district) were obtained from the HCMC Statistical Office.¹⁷ As population figures for those aged <16 years per ward or district was not obtainable, the total population was used to estimate disease incidence. Patient addresses in HCMC were geocoded to districts and wards. City level weekly average climate data during 2005–2010 were obtained from the Ministry of Natural Resources and Environment of Vietnam including relative humidity (median=76.5%, range=61.7%–88%), temperature (median=28.1°C, range=24.4°C–31.8°C), rainfall (median=24.4 mm, range=0 mm–230.7 mm) and water level of the Dong Dien River (median=4.93 cm, range=−34.43 cm to 49.43 cm).¹⁸

Spatial mapping

The variance of raw ward ARI incidence rates (number of ARI cases per 1000 ward total population for the whole study period) may largely vary due to the large variation in ward population size in HCMC. To take into account the spatial heterogeneity of variance and any dependence of ARI incidence rates between wards, ward ARI incidence rates were smoothed using local empirical Bayes estimates for rates reduced to a neighbourhood mean.¹⁹ The neighbourhoods were selected using the neighbourhood list based on wards with contiguous boundaries (R package spdep V.0.5–88).²⁰ Kulldorff and Nagarwalla’s method over the centroids was be used to scan for clusters of ARIs.²¹ This method allows the detection of possible cluster of any size, at any location in a population with inhomogeneous spatial density of HCMC. A Poisson generalised linear model was used to explore the relationship between district ARI incidence and district characteristics. CIs and tests were based on robust SEs to control for mild violation of the distributional assumption that the variance equals the mean.²²

Time series analysis

City-wide ARI weekly incidences (number of cases per 1000 population) were examined for seasonality, time trends and associations with climatic covariates by Poisson generalised additive mixed models (GAMMs) implemented in the R package mgcv V.1.8–7.^{23 24} The seasonal cycles and time trends of weekly ARI incidence were evaluated by the basic GAMM models containing the week of year (Ws=1–52) as cyclic cubic regression splines and the week of the whole study period (Wt=1–312) as cubic regression spline. As weekly ARI incidence may be associated with the ARI incidence and climatic condition of the previous weeks, in addition to the elements in the basic model, the full GAMM model contained lags of ARI weekly incidence up to 8 weeks, concurrent climatic condition and lags of climatic condition up to 8 weeks. The performance of the GAMM models was evaluated using adjusted R² and model diagnostic plots. Temporal correlation of the model residuals was evaluated by autocorrelation function and partial autocorrelation function.

All analyses and results were based on the records with available relevant information (complete cases). Spatial mapping and statistical analyses were stratified by URIs and LRIs and all analyses were performed in R V.3.2.2.²⁵

Basic features of ARIs in the paediatric population of HCMC

From 2005 to 2010, there were 155 999 children (<16 years old) residing in HCMC admitted to the three study hospitals for ARIs (33% of all admissions); 323 245 patients were admitted for conditions other than ARIs. There were 68 120 patients with ICD-10 codes synonymous with acute URIs (14% of all admissions) and 87 879 patients with ICD-10 codes associated with acute LRIs (19% of all admissions). The proportion of men (62%) and women (38%) was comparable between URIs and LRIs. The median age of the patients was 1.5 years (IQR 0.7–2.9 years). However, patients with LRIs were significantly younger than those with URIs (median age in years (IQR)=1.1 (0.5–2.1) vs 2.1 (1.1–3.9)), Wilcoxon’s test P<0.0001) and patients with LRI were hospitalised significantly longer than those with URIs (median days of hospitalisation (IQR)=6.0 (4–9) vs 4.0 (2–5), Wilcoxon’s test P<0.0001) (table 1).

Overall, the most common diagnoses were acute pharyngitis (28% of all ARIs), pneumonia (21% of all ARIs), bronchitis (18% of all ARIs) and bronchiolitis (16% of all ARIs). Among the patients with URI, the majority of patients were diagnosed with acute pharyngitis (63% of all URIs), acute URIs of multiple unspecified sites (14%), acute tonsillitis (8%), acute laryngitis and tracheitis (6%), and acute nasopharyngitis (6%). Among the patients with LRI, the most common diagnoses were pneumonia with an unspecified organism (37% of all LRIs), acute bronchitis (32%), acute bronchiolitis (29%) and bacterial pneumonia not elsewhere classified (2%) (table 1).

The spatial distribution of paediatric ARIs in HCMC

The raw ward-level incidence rates of ARIs for the entire study period ranged from 0 to 50 per 1000 total population (median (IQR)=18.8 (14.2–23.5)) and the smoothed ward-level empirical Bayesian estimated ARI incidence rates ranged from 1.2 to 48.8 ARI cases per 1000 population (median (IQR)=18.6 (14.7–23.7)). High incidence rates (>30 cases/1000 population) of ARI were observed in the wards surrounding the three study hospitals (wards in districts 1 and 4 in central HCMC) and in multiple wards located outside the central region of the city, specifically wards located in districts 2, 7 and Nha Be (observed vs expected incidence rate ratio (RR) >1.4, P<0.0001 using Kulldorff and Nagarwalla’s scan) (figure 1A). Notably, the spatial distribution of ARIs and diseases other than ARI were comparable (figure 1A,B). We observed an increase with time in ward incidence of ARIs in the majority of the wards with the most apparent high incidences in wards in the south of the city (online supplementary figure S1). We additionally found the spatial distribution of URIs and LRIs were comparable (online supplementary figure S2).

The district level incidence over the entire study period of ARIs was significantly lower in the districts located further away from the hospitals (mean district incidence=26.7 per 1000 population, RR for two times increase in distance to hospital=0.88, 95% CI 0.80 to 0.96) and in rural districts (RR for urban vs rural districts=1.54, 95% CI 1.10 to 2.15) (figure 1A). Other district characteristics were not significantly associated with district ARI incidence.

Temporal patterns and climatic risk factors of paediatric ARIs

The mean of weekly incidence of URIs for the study period was 3.12/1000 population. Weekly URI incidence showed a relatively small increase with time, with a difference in mean weekly URI incidence at the end and the start of the study period of 0.86 patients per 1000 population. The basic Poisson GAMM model indicated that URI admissions in the study hospitals exhibited a significant increasing trend with time (P<0.0013) and seasonality (P<0.0001) with a peak in May–June and a smaller peak in December–January (figure 2A).

The mean weekly LRI incidence for the study period was 4.02/1000 population. There was a significant increase of weekly LRI incidence over time (P=0.0037) with a substantial difference in the mean weekly LRI incidence at the end versus the start of the study period of 1.54 patients/1000 population. The seasonal cycle of LRIs (P<0.0001) was more obvious than that of URIs, with a single, wide annual peak in the months of September and October (figure 2B).

In a univariate analysis with Poisson generalised linear model, we found that weekly URI incidence was significantly associated with weekly average temperature (RR for 1° increase of temperature=1.048, 95% CI 1.030 to 1.066) and humidity (RR for 1% increase in relative humidity=1.006, 95% CI 1.001 to 1.011). Further, weekly LRI incidence was significantly associated with weekly average humidity (RR=1.030, 95% CI 1.024 to 1.036) and rainfall (RR for 1 mm increase in rainfall=1.019, 95% CI 1.014 to 1.024) (figure 3). However, after controlling for seasonality, time trend, lags of weekly incidence and other climatic covariates in the full GAMM model, the weekly URI incidence was significantly associated with 1-week and 2-week lags in URI incidence (RR=1.202, 95% CI 1.159 to 1.245 and RR=1.063, 95% CI 1.019 to 1.109, respectively), 1-week and 3-week lags in weekly average humidity (RR=1.009, 95% CI 1.003 to 1.014 and RR=0.990, 95% CI 0.984 to 0.996, respectively), a 3-week lag in weekly average rainfall (RR=1.006, 95% CI 1.003 to 1.009) and a 4-week lag in weekly average river water level (RR for 1 cm increase in water level=1.003, 95% CI 1.002 to 1.005) (table 2).

After adjusting for other factors in the full GAMM model, weekly LRI incidence was significantly associated with a 1-week and 2-week lags in LRI incidence (RR=1.173, 95% CI 1.138 to 1.208 and RR=1.053, 95% CI 1.015 to 1.092, respectively), a 1-week lag in weekly average temperature (RR=0.976, 95% CI 0.953 to 0.998), a 1-week lag in weekly average rainfall (RR=1.004, 95% CI 1.001 to 1.007) and a 1-week lag in weekly average water level (RR=0.998, 95% CI 0.997 to 1.000) (table 2). The time trends remained significant for both URIs (incidence RR for 1-week increase in time (RR)=1.09, 95% CI 1.02 to 1.17, P=0.0096) and LRIs (RR=1.08, 95% CI 1.006 to 1.16, P=0.0345) while only LRIs exhibited significant seasonality (P<0.0001) in the full GAMM models. The full GAMM models fitted the data well for both URIs (adjusted R²=0.72) and LRIs (adjusted R²=0.88) (figure 2 and online supplementary figure S3). Temporal correlation of weekly incidence was well controlled for URIs and also relatively well controlled for LRIs in the full GAMM models (online supplementary figure S4).