Identifying the role of household immunity in driving individual dengue virus infection risk

Kamphaeng-Phet family-based cohort study

This study used data from an ongoing family-based longitudinal cohort study in Kamphaeng-Phet, Thailand. Details of the design have been previously described ¹³. Briefly, we enrolled pregnant mothers and their multi-generational households. Per the inclusion criteria, a household was eligible for enrollment if a minimum of three members in addition to the newborn consented/assented to participation; the pregnant woman, another child, and an older adult aged at least 50 years. Active surveillance began with the birth of the newborn, with enrollment specimens for the remainder of the household collected prior to the birth of the newborn. In order to ascertain subclinical infections, serum samples were obtained from all participants roughly annually after enrollment. Acute febrile illness events were detected through a combination of active and passive surveillance strategies. Individuals were instructed to notify study staff if an acute febrile illness event occurred. In addition, participants were contacted by the study team on a weekly basis to determine if any member of the household had been febrile since the last contact. Upon discovery of a febrile episode, an illness investigation was triggered, where acute and convalescent blood samples were obtained from the febrile case. If the illness investigation identified a PCR-confirmed case a household investigation was triggered where acute and convalescent samples were taken for the remaining household members. The convalescent samples were taken at 14 and 28 days after the acute sample collection.

This study was approved by Thailand Ministry of Public Health Ethical Research Committee, Siriraj Ethics Committee on Research Involving Human Subjects, Institutional Review Board for the Protection of Human Subjects State University of New York Upstate Medical University, and Walter Reed Army Institute of Research Institutional Review Board (protocol #2119).

Laboratory methods

All samples obtained during routine visits were tested using hemagglutination inhibition assays (HAIs) to quantify antibody titers against all four DENV serotypes and Japanese encephalitis virus (JEV)¹⁴. Additionally, all acute and convalescent samples were tested by HAI for all four DENV serotypes and JEV as well as immunoglobulin M (IgM) and immunoglobulin G (IgG) capture ELISAs for DENV and JEV ¹⁵. All acute samples also underwent DENV reverse-transcriptase polymerase chain reaction (RT-PCR) ¹⁶. For the purpose of this analysis, we defined a confirmed DENV infection as any case that is RT-PCR positive for any DENV serotype or where both HAI and ELISA results using the acute and convalescent samples were diagnostic of an infection ¹².

Statistical analysis

The purpose of this analysis was to investigate individual and household risk factors for DENV infection in this multigenerational cohort. To do this, we first fit a classification algorithm to the yearly HAI data in order to identify subclinical infections. We then used these imputed infections to investigate individual and household-level drivers of infection.

Training data

We define a positive and negative person period as follows. A total of 90 confirmed DENV infections were identified through the case investigations. Data from the yearly HAIs surrounding these confirmed DENV infections were defined as the positive person periods. For negative person periods we took the remaining full dataset and removed any interval where an individual had a confirmed DENV infection via the illness investigation, or where individuals had a larger than fourfold increase in any one of their yearly DENV serotype HAIs. We then removed any individuals living in the same household during these aforementioned intervals since DENV transmission is known to be clustered within households. This left 3466 intervals that could potentially be used as negative controls from the available 11131 observed intervals (Figure S1). We randomly sampled a third of these to be added to the training data creating a total of 1246 intervals in our training set. The first interval of sampling for newborns was excluded in this analysis due to limited representation in the serologically supported infections that could provide information on maternal antibody kinetics.

Prediction model

Using training data described above, we ran a gradient boosted regression using the R package xgboost ^17,18. Unlike in random forest models where multiple independently trained decision trees are combined to determine the overall likelihood of a model, in gradient boosted regression each decision tree is fit on what the previous trained ensemble of trees have misclassified, allowing for refinement on difficult classification problems. The candidate predictors we used to train this model are listed in Table S3. Variables used to summarize the ratio and difference between pre and post-interval DENV titers across serotypes (maximum, minimum, geometric mean, and variance) were calculated at the individual serotype level and then the summary statistic of interest was quantified across all four serotypes.

Model fit

For hyperparameter tuning we utilized a random search approach within a nested cross validation approach where we initially split the training data into four cross-validation sets and subsequently performed hyperparameter tuning on each subset using five-fold cross validation. Model performance was quantified using the hold out set. Prior to each random search run we randomly downsampled the dataset to balance the number of positive and negative person periods. We performed this random search approach a total of 5000 times and saved the top 100 performing models evaluated on the held out cross validation set with the lowest log-loss value. The average predicted classification score (bounded between 0 and 1) for these 100 runs was taken to be the probability the individual was infected in that yearly interval. Intervals assigned a value greater than or equal to 0.5 were considered to be DENV infections.

Predicting subclinical DENV infections

We subsequently fit the models with the lowest log-loss values on the entire training dataset and predicted the presence or absence of infections in the remaining intervals that make up the evaluation data set. We used the training labels as ground truth and subsequently analyzed risk factors for the entire dataset.

Characterizing risk factors of DENV infection and disease

We fit a series of univariate and adjusted logistic regressions to characterize how DENV infection risk is a function of temporal, individual, and household factors. These models were run using the glmmTMB function found within the glmmTMB package in R ^19,20. We fit all models using a binomial GLM with a logit link function. All generalized linear models were optimized using the nlminb method found in the stats package. Only household random effects were incorporated into the model as the inclusion of both household and individual random effects led to singular fits.

Individual and household level risk

We first tested whether demographic factors were associated with risk including age, sex, and employment. We binned individuals into five age bins (1-5, 5-18, 18-30, 30-50, and 50+). Individuals under 1 year were excluded as they will usually have maternal antibodies, which would complicate this analysis due to different kinetics. Sex was defined upon enrollment into the cohort. Further information on individual and household related covariates can be found in the supplemental information.

We subsequently performed analyses to quantify how household composition and infection history impacted risk for an individual. Data on household composition consisted of the number of newborns, individuals between 1 and 5, individuals between 5 and 18, and adults all of which were broken down by sex. We fit models to estimate how the number of individuals in each of these bins impacted infection risk. For the analysis on infection history we took an individual of interest’s household and found the attack rate from the previous interval for all other members. We fit models to assess how the attack rate, the proportion of the household members who were inferred to have an infection, in the previous interval (categorized into three sets defined as containing strictly 0, (0.2), and [0.2,1]) impacted the infection risk. The distribution of these values was zero inflated and skewed right due to many households having no infections in the previous interval. Note that we removed the individual of interest in determining both the household composition and attack rate of the household in order to isolate how the household is impacting risk. For both of these covariates we fit three logistic regression models, a univariate model, a univariate model with random effects, and a multivariate model with random effects.

Since the goal of these models was to characterize the independent effect of the household-level covariates, each of these multivariate models also accounts for the individual’s average pre-interval HAI titer as well as the month and year of sampling as these have been shown to be important predictors of risk. This ensured the individual’s age, titers, and infection history did not impact subsequent calculations. Confounding effects of household related factors were accounted for in adjusted analyses where household random effects were incorporated. Note that as this analysis required at least two consecutive intervals, around 25% of the intervals were not included leaving 6913 intervals.

We then performed logistic regressions to understand how individual and household immunity impact DENV infection risk. We defined individual immunity to be the geometric mean of HAI titers transformed into log2 space. We defined household immunity for a particular individual to be the geometric mean of HAI titers of the household transformed into log2 space with the individual of interest removed from the calculation. HAI cutoffs of 40 and 66 were chosen for the household immunity covariate as these constituted the 33rd and 66th percentiles. Similar to the previous analysis, we fit three logistic regressions for each, a univariate model, a univariate model with random effects, and a multivariate model with random effects. In addition to these random effects and the covariate of interest, each multivariate model also accounts for the month and year of sampling. The household immunity adjusted model also accounted for the individual’s average pre-interval HAI titer.

Cohort description

The cohort study contained 3514 individuals within 515 households that have been enrolled since the study began in September 2015. 2868 individuals within 470 households were eligible for inclusion in this analysis as they completed at least one follow-up from enrollment through the cut-off for this analysis in late March 2022. The analyzed dataset contains data on 11131 “yearly” intervals, with an average of 3.90 intervals per enrolled individual (95% CI 1-6). Characteristics of the analyzed intervals are described in Table 1. These intervals were an average of 407.8 days long (95% CI 229 - 642.75) and took place over six sampling periods (Figure 1A). Not all individuals in a household consented/assented to being sampled, with approximately 80% of potential individuals sampled. Over the study period there were 469 index cases that triggered an illness investigation resulting in the identification of 107 infections, 90 of which occurred between paired yearly samples. These 90 infections consist of 62 PCR positive and with the remaining 28 being identified through serological evidence.

• Download figure
• Open in new tab

Figure 1.

Summary of cohort data. (A) Hospitalization counts for Kamphaeng Phet from 2015-2021 (blue solid line). Bars represent the timing of the confirmed DENV infections used to train the model. Serotype information was ascertained via RT-PCR. Shaded time periods represent active sampling periods during the cohort study when yearly blood draws were taken. (B) Age-stratified seropositive at enrollment for subjects enrolled before 2017. Points are mean seroprevalence found at each tenth percentile age bin. (C) Average DENV HAI titers at enrollment by age. Confidence bounds for B and C are found using a basic nonparametric bootstrap.

Model performance

We developed a classification algorithm that takes antibody titers measured during yearly visits to infer undetected infections in individuals. Our best fitting model was able to classify our training data with 93.3% sensitivity and 96.6% specificity (Figure S7). The longitudinal design of the cohort study allows for the visualization of HAI trajectories across time for enrolled individuals. Figure 2A illustrates imputed infections for three individuals enrolled in the cohort. The average and maximum ratios of pre to post interval HAI titers are the features of greatest importance for accurate classification defined by the information gain metric (Figure 2B).

• Download figure
• Open in new tab

Figure 2.

Summary of model performance and fits. (A) Three examples of HAI titer trajectories for all four DENV serotypes and JEV in three subjects. Alternating white and gray time periods represent distinct intervals, separated by the sampled HAIs. The imputed probability of an infection having occurred within an interval is presented by point size at the post-interval sample date while red arrows represent an imputed infection. Black arrows represent JEV vaccination events. (B) Feature importance across 100 model fits. Gain represents the relative contribution of each feature. (C) Pre and post interval HAI titers for the DENV serotype with the largest ratio grouped by age at post-interval sampling event and colored by whether the model predicted a seroconversion. Red points represent the confirmed seroconversions from the training data. A four-fold increase in titers between samples is represented by the black diagonal line.

Characterizing subclinical infections

Using our best fitting prediction models, we identified 1049 infections in 11131 intervals of observation, or 9.4% of intervals. This translates to 8.42 infections per 100 people per year (95% CI 7.71 - 9.15). This is similar to an estimated annual proportion of seronegative individuals being infected per year of 10.8% derived using a serocatalytic model from age-stratified seroprevalence data (95% CI, 9.9-11.8% using seropositive cutoff as HAI >=20) OR (8.0%, 95% CI, 7.3-8.7% using seropositive cutoff as HAI >=40) (Figure 1B). This force of infection is derived for non-newborns under 30 years old and the methods taken to estimate it are outlined in the supplemental information. We note that the model had high certainty in the assigning of infections for the majority of infections, with 673 of the 1049 intervals with infection being given a probability of greater than 90%. Similarly, the model had high certainty for the absence of infections in the remaining intervals with 8455 of the 10285 intervals being assigned a probability of less than 10% (Figure S8).

We found that the incidence of infections varied by year, with 2018 having higher incidence (Figure 3A). This is the same year where hospitalizations peaked in Kamphaeng Phet across the active years of the study (Figure 1A). Incidence of infection rates peaked amongst school aged children while individuals over the age of 30 all experienced lower incidences (Figure 3B). As expected, the proportion of primary infections (infections occurring in individuals without detectable antibodies to any serotype in any prior visit) was directly related to age, with almost all infections being post-primary (occurring in individuals with HAI antibody titers against at least one serotype greater than 20) after age 25. The ratio of subclinical to symptomatic infections was 13.8:1 (95% CI: 10.0-17.8:1) in the cohort. There was some variability across years and age, with the highest risk of symptomatic disease occurring between the ages of 15-25 (Figure 3D-E). We note that there were only 77 symptomatic infections out of 1049 total infections, leading to wide confidence intervals for these ratios, particularly for years of age-groups with few cases. Out of these 1049 infections, 139 individuals had multiple infection events throughout the study. The probability of having a second or third infection given that the individual had a previous infection peaks between the ages of 10-15, similar to the age range of highest incidence (Figure S2B).

• Download figure
• Open in new tab

Figure 3.

Incidence, proportions of primary infections, and subclinical to symptomatic ratios across time and age. (A-B) Incidence (infections per person-year) for both symptomatic (red) and subclinical (yellow) infections across interval year (A) and age (B). (C) Proportion of primary infections as a function of age. (D-E) ratio of subclinical to symptomatic DENV incidence in the cohort as a function of interval year (D) and age (E). Mean and 95% CIs for the ratio of subclinical to symptomatic DENV incidence are represented by the dotted line and gray regions respectively. Mean and 95% CIs for polynomial fits to time and age are represented by the solid blue line and regions.

Risk factors for DENV infection

Using the imputed infections from the classification algorithm we investigated which individual and household risk factors were associated with infection risk. When compared to young children between the ages of one and five years old we found that individuals between 5 and 18 as well as those between 18 and 30 were at higher risk of infection with an aOR of 1.44 (95% CI 1.16-1.77) and 1.41 (95% CI 1.06-1.89) respectively. In an unadjusted analysis there was no significant difference in odds of infection by sex when comparing males to females (OR 1.11, 95% CI 0.98 - 1.27). However, our data is consistent with an observed interaction between age and sex in infection risk with women between the ages of 18-40 being at a higher risk of infection than their male counterparts (Figure S4). We also studied how an individual’s occupation may impact their risk and found no differences that attained statistical significance in the adjusted analysis (Table S1).

We subsequently studied how household level factors impact an individual’s risk of infection. No covariates describing the surrounding built environment had a significant impact on dengue risk. However, we did find strong associations between household composition and risk of infection. While the number of individuals living in the household was not associated with risk of infection (aOR 1.00, 95% CI, 0.97 - 1.04), we found that each additional adult in the household reduced the likelihood of infection in the other household members with an aOR of 0.95 (95% CI, 0.90 - 0.99). The presence of each additional newborn and individual between 5 and 18 increased the odds of infection for the other household members with an aOR of 2.13 (95% CI, 1.65 2.75) and 1.09 (95% CI, 1.01 - 1.19) respectively. Although not significant, the presence of each additional individual between one and five increased the odds of infection for household members with an aOR of 1.13 (95% CI, 1.00 - 1.28) (Figure 5A). Analyses stratified by sex revealed a more complex association between household composition and risk. For either sex, each additional newborn increased infection risk for the other individuals living in that household. For older age groups however the associations varied by sex. Each additional male between the ages of 1 and 5 and between 5 and 18 increased risk with an aOR of 1.25 (95% CI, 1.08 - 1.44) and 1.18 (95% CI, 1.06 - 1.31) respectively while additional adult male had no impact on risk. Additional females provided no changes in risk except for adults, where each additional female adult reduced risk with an aOR of 0.88 (95% CI, 0.81 - 0.95) (Figure 5B).

Beyond characterizing the association between household characteristics and composition on dengue risk, we sought to understand the impact of individual and household immunity. Consistent with previous findings, the most important predictor of infection risk during an interval was an individual’s HAI titers at the beginning of the interval (Figure 4). In this analysis, the magnitude of average HAI log2 titers was inversely associated with risks of both symptomatic and subclinical infections. On average, each log2 increase in titers was associated with a 26.3% (95% CI: 23.3 - 29.1%) decrease in risk of infection and a 38.7% (95% CI: 27.9 - 47.9%) decrease in having a symptomatic infection. Interestingly, we also found that household immunity impacted an individual’s risk of infection even when accounting for that individual’s antibody titer. The distribution of these variables is found in Figure S7. Individuals living in households with high immunity (average HAI titers greater than 66) had decreased risk with an aOR of 0.78 (95% CI, 0.63 - 0.96) when compared to those with an average below 40 (Figure 5D). Since household titers are likely to be associated with recent household infection history, we also investigated how household attack rates during a preceding interval (the proportion of individuals within a household that had an imputed DENV infection in the preceding interval) impact future risk. Individuals living in households that had moderate to high attack rates (greater than 20% of household members experiencing an infection) during the previous year were at decreased risk of infection with an aOR of 0.61 (95% aOR CI, 0.49 - 0.77) in comparison to individuals coming from a household with no infections the previous year (Figure 5C).

• Download figure
• Open in new tab

Figure 4.

Geometric mean of pre-interval DENV titers across all four serotypes and their impact on probability of infection and symptoms. We first calculated the annualized probability of infection and then fit splines of order three to the data. (A) Probability an individual is infected in a year as a function of their pre-interval DENV titers. (B) Probability an individual is symptomatic in a year as a function of their pre-interval DENV titers. (C) Probability that an individual is symptomatic given that they were infected in the same year as a function of their pre-interval DENV titers.

• Download figure
• Open in new tab

Figure 5.

Impact of household composition (A-B), infection history (C) and immunity (D) on risk of infection. (A) Odds ratio for the number of total individuals in various age bins (newborn [NB], from 1-5 years old [LT5], from 5 to 18, and those older than 18 [GE18]) defined at the time of the post-interval sample. (B) Odds ratio for the number of males and females of various age bins (newborn [NB], from 1-5 years old [LT5], from 5 to 18, and those older than 18 [GE18]) defined at the time of the post-interval sample. (C) Previous interval’s attack rate and subsequent odds ratio of infection risk relative to having no infections in the previous interval. (D) Geometric mean of DENV HAI titers for the rest of the household members and subsequent odds ratio of infection risk relative to having an average household HAI titer under 40. All models are adjusted for household random effects, individual pre-interval titers, as well as the year and month of post-interval sample.