HIV Infection as Risk Factor for Death among Hospitalized Persons with Candidemia, South Africa, 2012–2017

HIV-seropositive persons demonstrated increased adjusted risk for 30-day mortality and should be evaluated for intensive care.

M60 Candida species-specific breakpoints for MICs (MIC); this was restricted to species or antifungal agents for which breakpoints have been published (5). Inappropriate antifungal treatment was defined as treatment around the time of diagnosis with an antifungal agent to which the Candida species was resistant. We defined a mixed episode of candidemia as >1 Candida spp. either cultured from a single blood culture specimen or from multiple specimens over the 30-day period. We defined community-onset infection as a positive blood culture within 72 hours of hospital admission. We considered age and sex to be important a priori confounders.
We screened potential confounders by cross-tabulating these variables with the main exposure and looked for associations in our dataset. None of the potential confounders were considered to be on the causal pathway for the main study question. We decided that the following potential confounding variables had collinear relationships with the main exposure: ICU admission, central venous catheter in situ, and total parenteral nutrition. We chose to use ICU admission in our models.

Classical Analyses
We looked for evidence that the effect estimate for the association between HIV status and 30-day mortality was confounded or modified by each explanatory variable. We compared the summary weighted odds ratio (OR) to the crude OR and looked for a ≈10% change in main exposure effect estimate as evidence of confounding. We also visually compared the stratumspecific ORs across each level of the potential effect-modifying variable and performed a test of homogeneity of ORs.

Model Building
We omitted potential confounders with a large amount of missing data. We checked the number of events and then applied a rule of ten (events per variable) to ensure that the models included a reasonable number of parameters. We forced the prespecified confounder variables (i.e., age, sex) into a baseline model and then added other potential confounders, one at a time using a forward approach, starting with the variable with the strongest association with death on univariable analysis. We retained a potential confounder in the model if this changed the main effect estimate by 10% and then added the other variables in turn. We checked the reliability of parameter estimates by using the quadchk command in Stata (https://www.stata.com) because the number of participants within many of the 29 clusters was large (>20). We compared the final logistic regression models with and without an interaction term between HIV status and ICU admission.

Description of 1,040 Cases Included in Analysis
The number of sentinel sites and thus the number of cases varied by year of surveillance: All sites were public-sector hospitals except 3 private hospital sites that participated in 2016 and 2017. We conducted a complete records analysis. Among 1,040 cases included in this analysis, 730 (70%) case report forms were completed by interview plus medical chart review versus 310 (30%) by retrospective medical chart review alone. In general, the completeness of case report forms was better for participants who were interviewed and had their medical charts prospectively reviewed. Among the 806 cases that were excluded because of missing HIV status or outcome data, 374 (46%) had case report forms completed by interview or medical chart review versus 432 (54%) by medical chart review alone (p<0.001). The median time of survival among 307 participants who did not receive systemic antifungal treatment was 3 days (interquartile range [IQR] 1-10 days) from the positive blood culture date versus 16 days (IQR 8-32 days) among 686 who received treatment.

Unadjusted Risk Factors for 30-Day Mortality
The case fatality ratio was incrementally higher in each of the 3 older age groups compared with the baseline category (i.e., 18 months-17 years) (p<0.001) (

Classical Stratified Analysis of the Relationship between HIV Status and 30-Day Mortality
Upon classical stratified analysis, the summary odds of death at 30 days remained approximately 2-fold higher among HIV-seropositive participants versus HIV-seronegative participants with candidemia when adjusted for each potential confounder in turn. We found consistently strong evidence against the null hypothesis of an OR = 1 (p<0.001 in all cases, except when adjusting for inappropriate antifungal treatment) (Appendix Table 2 when we adjusted for inappropriate antifungal treatment. However, as already noted, we had a large amount of missing data for both the latter variables. Using a low-power test for interaction, we did not find strong evidence of an interaction between HIV status and any other variable when comparing 30-day mortality rates among HIVseropositive persons with those of HIV-seronegative persons. Conversely, upon visual inspection of the stratum-specific rate ratios in Appendix Table 2, we identified the following as possible effect modifiers (different stratum-specific OR point estimates but overlapping 95% CI): age, sex, year of diagnosis, ICU admission, inappropriate antifungal treatment, and removal of the CVC. However, we focused only on ICU admission as an effect modifier because we had specified this a priori.

Random-Effects Multivariable Logistic Regression Analysis
We conducted a 2-level random-effects multivariable logistic regression analysis. We excluded 133 cases with missing data for any of the confounders; thus, the final model included 907 participants. We excluded quick Pitt score and inappropriate antifungal treatment from the final analysis because there were large amounts of missing data, even though these were potentially important confounders. However, when these variables were retained in a model with fewer observations (n = 263), the point estimate for the effect of HIV infection on mortality rates was very similar to that described in the final model (Appendix Table 3).

Supplementary Discussion
In 2 small studies of HIV-seropositive participants with candidemia, infection with a species other than C. albicans was independently associated with death (6,7). On the basis of these previous reports, we initially hypothesized that HIV-seropositive persons would be more likely to be infected with a species other than C. albicans. Since these species are more likely to be antifungal-resistant, we then speculated that death would be more likely in this group because of inappropriate antifungal treatment. However, we found that HIV-seropositive participants had a 34% lower odds of being infected with Candida species other than C. albicans. This is also consistent with our finding that almost a third of HIV-seropositive persons may have acquired candidemia as a community-onset infection, though a similarly large proportion of HIVseronegative persons also had a diagnosis within 72 hours of admission. We did not collect information on all healthcare contacts in the preceding 12 months so could not exclude that these infections were healthcare-associated. Community-onset infections in HIV-seropositive persons have been documented in a single-center study in South Africa (8). Despite its susceptibility to most antifungal agents, C. albicans, a common component of the gastrointestinal microbiota, has a well-described array of virulence factors that may increase severity of illness and risk for death (9). In a study in Italy, persons with candidemia caused by >1 Candida spp. versus monomicrobial infection were more likely to be HIV-seropositive (2/15 vs. 14/737). In turn, mixed candidemia was independently associated with increased deaths; the authors hypothesized that this effect was mediated by a larger fungal bloodstream burden in mixed infections (10). In our study, only 3% of patients had a mixed infection, and proportions were similar in HIVseropositive groups and HIV-seronegative groups.  (1), systolic blood pressure <90 mm Hg (1), cardiac arrest (1), mechanical ventilation (1), and altered mental status (1) at time of diagnosis of candidemia.