Impact of asymptomatic Plasmodium falciparum infection on the risk of subsequent symptomatic malaria in a longitudinal cohort in Kenya

Essential revisions:

1. How was the cohort assembled to support the representativeness of the participants for the general community?

The cohort was assembled using radial sampling of 12 households per village for three villages within Webuye, Western Kenya. The first household in each village was randomly selected. The three villages were geographically close and chosen due to their high malaria prevalence in a previous cross-sectional study in the area [Mean P. falciparum prevalence in 2013: Kinesamo (18.4%), Maruti (20.8%), Sitabicha (22.8%), Obala AA PloS One 2015;10(7)]. We have added some text to the methods to describe this:

“The cohort was assembled using radial sampling of 12 households per village for three villages with high malaria transmission. The first household in each village was randomly selected. Two households moved during follow-up and were replaced.”

2. Which symptoms were asked about in the self-report, and what was the algorithm used to define a case as symptomatic or asymptomatic, over what time period?

We have clarified this by adding to the Exposure and outcome ascertainment section of the manuscript: “The main exposure was an asymptomatic P. falciparum infection during monthly active case detection assessments, defined as P. falciparum-positive by qPCR in a person lacking symptoms. People who were P. falciparum-negative by qPCR during monthly visits were considered uninfected. We defined symptomatic P. falciparum infection as the current presence of at least one symptom consistent with malaria during a sick visit (i.e. fever, aches, vomiting, diarrhea, chills, cough or congestion) and P. falciparum-positive by both RDT and qPCR.” Additional details surrounding exposure and outcome ascertainment are provided in the supplement as described in #3 below.

3. Was any participant classified as symptomatically infected at a monthly visit? If so what status were they given at that visit?

Yes, and this information is now described in the supplement under the header Exposure and outcome ascertainment:

“Some participants were classified as symptomatically infected at a monthly visit through passive detection of symptoms; this occurred if a study team member conducting a monthly visit was approached by a participant reporting malaria-like symptoms. […] If that person did not meet our case definition for symptomatic malaria, then they were removed from follow-up for that month and re-entered for follow-up in the following month.”

4. How many participants were symptomatic, PCR positive but RDT negative, and therefore not treated? How did their clinical course, if any, compare with those who were also RDT positive and therefore treated?

In the primary analysis data set, we included 0 participants who were symptomatic, PCR positive, and RDT negative, because they did not meet our case definition for symptomatic malaria. Looking back at the original data set prior to any censoring criteria being applied, there were 6016 observations collected through monthly and sick visits and 183/6016 (3.0%) of them met the criteria of being symptomatic, PCR positive, and RDT negative. Subsetting the data set to only sick visits, 183/983 (18.6%) of sick visits were symptomatic, PCR positive, and RDT negative. Because this number was fairly few across a large number of participants and 29 months of observation, we have limited ability to make inferences about the natural history of malaria in RDT-negative/PCR-positive people; however, we did conduct a secondary analysis where we computed the hazard of symptomatic malaria using a secondary permissive definition for symptomatic malaria (PCR-positive and having at least one malaria-like symptom) to assess if alternative symptomatic malaria case definitions influenced results. Using the secondary permissive case definition for symptomatic malaria, we observed overall similar results to those produced using the primary case definition. We have added this to the main text:

“We observed similar 1-month elevated risks of malaria in asymptomatically-infected people when using both the “permissive” [aHR 1.97, 95% CI 1.63 to 2.40] and the “stringent” [aHR 2.76, 95% CI 2.11 to 3.62] alternate case definitions for symptomatic malaria (Figure 3—figure supplement 3).”

5. Was there any correlation by individuals in the same household?

We chose to only include a random intercept at the individual and not household level due to the little impact correlation at the household level had on the relationship between asymptomatic/uninfected at monthly visits and subsequent symptomatic malaria. We assessed the possibility of using a nested frailty approach with random intercepts at the individual and household levels. To do so, we re-computed the frailty Cox proportional hazards model described in Equation 1 in the main text using a random intercept at the individual level and an additional random intercept at the household level. Results were similar for the 1-month hazard of symptomatic malaria for those asymptomatically-infected versus uninfected when a random intercept at the household level was included [aHR: 2.64, 95% CI: 2.07 to 3.37] versus excluded [aHR: 2.61, 95% CI: 2.05 to 3.33]. Because including a household level random intercept had little impact on model results, we excluded it from analyses to be able to have model convergence to test effect measure modification of the relationship by age and sex.

6. Are you able to control for spatial heterogeneity in risk, that could also help tease out whether superinfection has confounded your results?

We controlled for spatial heterogeneity in risk by including village fixed effects in the models. We do not expect spatial risk to be substantially different within and between villages due to the radial sampling of households, close proximity of villages to each other (all less than 11 km apart), and similar high malaria prevalence across the three villages [Mean P. falciparum prevalence in 2013: Kinesamo (18.4%), Maruti (20.8%), Sitabicha (22.8%), Obala AA PloS One 2015;10(7)].

7. Please reconsider your suggestion of causality between asymptomatic and symptomatic infections.

We have relaxed the language in the main text to focus more on the association between asymptomatic and symptomatic infections instead of a causal relationship. To do so, we changed the casual language. For example, the Abstract now reads “Compared to being uninfected, asymptomatic infections were associated with an increased 1-month likelihood of symptomatic malaria [adjusted Hazard Ratio (HR):2.61, 95%CI:2.05–3.33]…” We made similar changes in the Introduction, Results, and Discussion, each of which are tracked in the accompanying documents.

8. Were the time periods that occurred after an individual was treated for a symptomatic illness different than before they had an illness? I noted that treatment was not included in the models.

To test if prior treatment during the study period influenced results, we re-computed the frailty Cox proportional hazards model in Equation 1 with a covariate for prior treatment. To do so, we reran the frailty Cox proportional hazards model in Equation 1 with an additional covariate representing prior receipt of antimalarial treatment. This variable was coded dichotomously as having received study-prescribed antimalarials up until that monthly visit or not. For example, a person was coded as having not received study-prescribed antimalarials up until their first symptomatic infection, but afterward were coded as receiving treatment from that point forward in follow-up. Compared to primary models, antimalarial treatment had minimal effect on the hazard of symptomatic malaria when asymptomatically infected versus uninfected in the short-term [1-month aHR: 2.61, 95% CI: 2.05 to 3.33] but made results largely null in the long-term [29-month aHR: 1.04, 95% CI: 0.95 to 1.15].

We have added to the methods section Sensitivity analyses:

“For prior antimalarial treatment, we included a variable coded dichotomously as having received study-prescribed antimalarials up until that monthly visit or not; a person was coded as having not received study-prescribed antimalarials up until their first symptomatic infection, but afterward were coded as receiving treatment from that point forward in follow-up,” and the Results the general statement that “This association was similar in a model adjusted for covariates [adjusted HR (aHR): 2.61, 95% CI: 2.05 to 3.33] (Table 2, Figure 3A) as well as when using alternative modeling approaches, alternate outcome case definitions, and in sensitivity analyses.”

9. Given that some parasite genotyping data are already available for this cohort, could the authors incorporate genotyping data in their analysis to distinguish between persisting infections that became symptomatic vs. superinfection with new strains?

Unfortunately, parasite genotype data are unavailable for more than half of the months of observation that we analyze here. Our recent analysis of parasite genotypes used data from the first phase of this cohort [Sumner KM Clin Infect Dis 2021;ciab357], and we supplemented these 14 months of phase 1 with an additional 15 months from phase 2 for this analysis.

Results from that paper [Sumner KM Clin Infect Dis 2021;ciab357] suggested an association between the presence of new haplotypes in incident infections and an increased likelihood of symptomatic malaria; however, this relationship was overall null for persistent infections. We have referenced some of these results into the discussion:

“The increased short-term hazard could reflect misclassification of a “pre-symptomatic” infection that progressed to symptoms as an asymptomatic exposure (Njama-Meya et al., 2004). […] The increased hazard of symptomatic malaria could also have been due to the presence of new genotypes in infections (i.e. superinfection), although we previously reported that such newly-apparent genotypes were associated with symptoms only in previously-uninfected people (Sumner et al., 2021).”

10. The authors ascertain the exposure status at every monthly follow-up visit, thereby treating each of these visits as a new study entry. It remains unclear whether this method is applicable for infectious diseases for which the risk of subsequent infections changes over time. The key question is whether this change is moderate or not over the time period considered. The stratification by age shows substantial changes over five year periods which is not that much more than a 29-month follow up. Please discuss whether these results are robust against the deviations from moderate time dependence.

To account for the gradual acquisition of immunity in cohort participants and subsequently time dependence, we have re-computed the models including a variable for the number of prior asymptomatic infections a person experienced in the study up until each monthly visit. To assess how the number of prior asymptomatic infections a person had during the study period could have influenced results, we re-computed the frailty Cox proportional hazards model in Equation 1 with a variable added for the number of prior asymptomatic infections each person had in the study up until each monthly follow-up visit. The covariate for number of prior infections was included as a continuous number in the model. We observed that the number of prior asymptomatic infections had little impact on the hazard of symptomatic malaria comparing those asymptomatically infected to uninfected in both the short-term [1-month aHR: 2.60, 95% CI: 2.03 to 3.31] and long-term [29-month aHR: 1.29, 95% CI: 1.17 to 1.42] compared to the primary models” Overall, we did not observe a large difference in the short-term or long-term results when accounting for the number of prior infections a person experienced, suggesting that the results were robust to deviations from time dependence.

We have added to the methods section Sensitivity analyses:

“For the number of prior infections, we included in the model as a covariate the number of prior infections as a continuous number,” and to the Results the general statement that “This association was similar in a model adjusted for covariates [adjusted HR (aHR): 2.61, 95% CI: 2.05 to 3.33] (Table 2, Figure 3A) as well as when using alternative modeling approaches, alternate outcome case definitions, and in sensitivity analyses.”

Reviewer #1:

[…] Comments for the authors:

This is a clearly written manuscript that addresses a topic of major public health significance. My only concern refers to the way the interesting results reported here are explored.

The authors use a relatively new approach to survival analysis to show that carriers of asymptomatic malaria infections are more likely than uninfected individuals in the same populations to develop a clinical disease over the next month, although not necessarily over longer periods of follow-up. To this end, they excluded putatively "pre-symptomatic" infections, that were arbitrarily defined as those occurring within 14 days prior to a symptomatic infection. I see no clear justification for this 14-day time window.

I would favor an alternative, more strict approach: excluding infections leading to clinical symptoms only if identical parasite haplotypes persisted since the first parasite detection until the clinical episode. This would provide evidence of "pre-symptomatic" infections. Combining changes in parasitemia over time would make the case even stronger. Quantifying the proportion of subsequent symptomatic infections associated with persisting haplotypes would provide an information of major biological and epidemiological interest and would allow us to test whether the arbitrary time window of 14 days is appropriate to rule out "pre-symptomatic infections". The authors have already genotyped a large proportion of incident infections in cohort participants (doi: 10.1093/cid/ciab357) but do not consider this rich dataset in their analyses.

We have updated the manuscript text to reflect why the 14 day (2 week) time window was chosen for the pre-symptomatic subset analysis:

“The time frame for identifying potentially pre-symptomatic infections was chosen for consistency with previous work studying time to symptomatic malaria (Buchwald et al., 2019).”

Unfortunately parasite genotype data are only available for less than half of the time period of this cohort. Additionally, incorporation of these complex deep-sequenced genotype data would vastly expand the scope and impact of this report, which is focused on the clinically-actionable aspects of the natural history of asymptomatic infections.

The exposure status of each study participant (whether asymptomatically infected or not) was ascertained at every monthly follow-up visit and allowed to vary each month. This is an interesting approach that has been previously used in epidemiological studies of non-communicable diseases, but not of infections. The caveat here is that each new infection (either persisting or self-cured) may reduce the subsequent risk of malaria-related disease. Therefore, someone who is uninfected (at baseline), then infected (first follow-up visit) and again uninfected (second follow-up visit) has a similar "exposure status" during the second follow-up visit as someone who remained uninfected since the baseline evaluation, but this (mis)classification ignores acquired immunity gradually developed in response to each infection experienced by cohort participants, which may reduce the subsequent risk of infection and disease.

Addressed in Essential revisions #10 above.

Not surprisingly, the risk of disease following asymptomatic parasite carriage increases with increasing parasite density (Figure 4), but I wonder whether age-related fever thresholds may be observed in this population. Given that parasite density data are available, I suggest that a further analysis of parasite density levels associated with symptoms across age groups might provide some interesting insights into the biological/immunological mechanisms underlying the main study finding.

As suggested, we did an additional analysis assessing the 1-month hazard of symptomatic malaria comparing people with asymptomatic infections with parasite densities above a series of thresholds to people who were uninfected, stratified by people’s age groups (<5 years, 5-15 years, and >15 years). Among adults >15 years, we still saw a positive association between increasing parasite density in asymptomatic infections and a higher 1-month hazard of symptomatic malaria (see Figure 4—figure supplement 1); however, parasite density had less of an effect among children. We have added this analysis to the main text:

“As an additional analysis, we repeated this process for each parasite density threshold stratified by participant age (<5 years, 5-15 years, >15 years). This observed increase in the hazard of symptomatic malaria with increasing parasite density was most pronounced among adults >15 years (Figure 4—figure supplement 1); however, children’s likelihood of symptomatic infection did not appear to be influenced by parasite density.”

Reviewer #2:

1) Seasonality was not considered in the analysis, but might be important to understand the observations. For example, the finding that "a difference in the time to symptomatic malaria in the first few months post asymptomatic infection but not long-term" could be confounded by the fact that most transmission occurs in the wet season, and thus that the probability for both asymptomatic and clinical infection is higher in the wet season.

To investigate the effect of seasonality on our results, we re-computed the models including seasonality as a covariate. To investigate how seasonality could have affected results, we re-computed the frailty Cox proportional hazards model in Equation 1 while including a variable for seasonality. We classified monthly visits that occurred any time from May to October as the high transmission season and from November to April as the low transmission season, based on the region’s rainy seasons. With seasonality in the model, a high hazard of symptomatic malaria was observed among those that were asymptomatically infected compared to uninfected at monthly visits [1-month aHR: 2.46, 95% CI: 1.93 to 3.15; 29-month aHR: 1.14, 95% CI: 1.04 to 1.26]; these results were similar to the main analysis presented in the manuscript using Equation 1, suggesting that the increased hazard of symptomatic malaria following asymptomatic infection was not solely attributable to seasonality.

We have added to the methods section Sensitivity analyses:

“For seasonality, we classified monthly visits that occurred any time from May to October as the high transmission season and from November to April as the low transmission season, based on the region’s rainy seasons.,” and to the Results the general statement that “This association was similar in a model adjusted for covariates [adjusted HR (aHR): 2.61, 95% CI: 2.05 to 3.33] (Table 2, Figure 3A) as well as when using alternative modeling approaches, alternate outcome case definitions, and in sensitivity analyses.”

2) Lines 52-54: It would be good to mention whether these estimates are based on microscopy or PCR. It would also be good to introduce the concept of supatent infections in the introduction, and discuss the possibility that even by PCR some infections were missed.

We have now added the specific malaria diagnostics used for the estimates:

“In 2015, a geo-spatial meta-analysis estimated a continent-wide prevalence of asymptomatic P. falciparum in children aged 2-10 years old of 24% based on microscopy and rapid diagnostic test results (RDT) (Snow et al., 2017).”

3) Lines 188-189: While I like brevity, this section should be expanded, given its importance for the study. How many participants were never infected? How often was someone infected at a single time point? What was the duration of infections? Did you observe the sequence infected-non infected-infected, pointing to possible undetected infections at certain time points (in particular if the same clone was observed)?

We have added some details to the first paragraph of results describing participant follow-up: “… person-months of asymptomatic malaria exposure; the median total months of asymptomatic exposure for a participant was 9 (IQR: 5, 17). Exposure status frequently changed for participants and remained constant for only 16 (6.2%) people across follow-up; 4 people were asymptomatically infected for the entirety of follow-up and only 12 people were never infected (Figure 2A).”

Likewise, given that age was an important variable, please provide more details on the age of clinical patients. A figure showing age trends in asymptomatic and clinical infections would be helpful.

We have added more detail about the age of the participants to the main text and have added Figure 3—figure supplement 2 illustrating the age distribution: “… and a median age of 13 years (range: 1, 85) (Figure 3—figure supplement 2).”

Reviewer #3:

Comments for the authors:

The paper is very well written and conceived, but I do not agree that the main conclusion is supported by the data.

Comments on the statistical analysis:

1. Line 105-107: missed monthly visits got an exposure status equal to the previous month's value – was any sensitivity analysis done?

Our primary models impute missed monthly visits to have an exposure status equal to the previous month’s value. We chose this approach based on a previous paper that had employed this approach to study asymptomatic malaria over time [Nguyen Lancet Infect Dis 2018;18(5)]. Imputing missed monthly visits added 715 observations to the primary analysis data set, with imputed monthly visits making up approximately 13% of the observations in the data set (715/5379).

We have added a sensitivity analysis to the manuscript where we repeated the 1 and 29-month hazard of symptomatic malaria analyses using the original data set without imputation for missed monthly visits. We found that results produced using the data set without imputation [1-month aHR: 2.75, 95% CI: 2.05 to 3.66; 29-month aHR: 1.40, 95% CI: 1.22 to 1.61] were similar to the results presented in the manuscript where imputation was performed [1-month aHR: 2.61, 95% CI: 2.05 to 3.33; 29-month aHR: 1.11, 95% CI: 1.01 to 1.22]; notably, imputation produced estimates closer to the null of 1, suggesting that bias from imputation was towards the null.

We have added to the Exposure and outcome ascertainment section “A sensitivity analysis was conducted for imputation using a dataset without imputation for missed monthly visits,” and to the Results “This association was similar in a model adjusted for covariates [adjusted HR (aHR): 2.61, 95% CI: 2.05 to 3.33] (Table 2, Figure 3A) as well as when using alternative modeling approaches, alternate outcome case definitions, and in sensitivity analyses.”

2. Line 107-109: why are people considered lost to follow-up and censored at the time of the imputed monthly visit; I would argue imputation is best not done in this situation.

Addressed above.

3. Line 136-138: How do the authors account for repeated measures using a Bonferroni correction?

We applied the Bonferroni correction to all the table p-values for Table 1 by multiplying each p-value by 29, which was the maximum amount of follow-up visits (and repeated measures) each person could have in the study.

4. Line 143-144; Equation 1: What is ${ϵ}_i$ in this formula? What are the underlying distributional assumptions for ${\alpha }_i$ and were these assumptions challenged? Please clarify notation used.

We have clarified the definitions for these terms in the main text:

“We allowed the main exposure to vary each month based on the monthly follow-up visit infection status ($m$), and included a random intercept at the participant level (${\alpha }_i$) to account for potential correlated intra-individual outcomes. A log-normal distribution was used for the random effect ${ϵ}_i$. represented the model’s error term.”

5. Individuals are clustered in households (38 in total); was this taken into account in the analysis? This would be very relevant for malaria. A nested frailty approach would be useful here.

Addressed in Essential revisions #5 above.

6. Throughout the text it would help to clarify which HR are referred to: conditional or unconditional (wrt the frailty). This can be briefly mentioned so that repetition is not needed.

Throughout the main text and supplement, we have updated the HR abbreviation to indicate whether the unconditional, crude HR is being reported (cHR) or the conditional, adjusted HR is being reported (aHR). Models that produced an aHR always included a random intercept at the individual level.