A Longitudinal, Observational Study of Etiology and Long-Term Outcomes of Sepsis in Malawi Revealing the Key Role of Disseminated Tuberculosis

Abstract Background Sepsis protocols in sub-Saharan Africa are typically extrapolated from high-income settings, yet sepsis in sub-Saharan Africa is likely caused by distinct pathogens and may require novel treatment strategies. Data to guide such strategies are lacking. We aimed to define causes and modifiable factors associated with sepsis outcomes in Blantyre, Malawi, in order to inform the design of treatment strategies tailored to sub-Saharan Africa. Methods We recruited 225 adults who met a sepsis case definition defined by fever and organ dysfunction in an observational cohort study at a single tertiary center. Etiology was defined using culture, antigen detection, serology, and polymerase chain reaction. The effect of treatment on 28-day outcomes was assessed using Bayesian logistic regression. Results There were 143 of 213 (67%) participants living with human immunodeficiency virus (HIV). We identified a diagnosis in 145 of 225 (64%) participants, most commonly tuberculosis (TB; 34%) followed by invasive bacterial infections (17%), arboviral infections (13%), and malaria (9%). TB was associated with HIV infection, whereas malaria and arboviruses with the absence of HIV infection. Antituberculous chemotherapy was associated with survival (adjusted odds ratio for 28-day death, 0.17; 95% credible interval, 0.05–0.49 for receipt of antituberculous therapy). Of those with confirmed etiology, 83% received the broad-spectrum antibacterial ceftriaxone, but it would be expected to be active in only 24%. Conclusions Sepsis in Blantyre, Malawi, is caused by a range of pathogens; the majority are not susceptible to the broad-spectrum antibacterials that most patients receive. HIV status is a key determinant of etiology. Novel antimicrobial strategies for sepsis tailored to sub-Saharan Africa, including consideration of empiric antituberculous therapy in individuals living with HIV, should be developed and trialed.


2
One aerobic blood culture was inoculated with 5-10ml of blood and incubated using an automated system (BacT/Alert, BioMerieux) in the MLW labs and cultured isolates identified to genus level (all) and species level using the API system and standard techniques 1 . Antimicrobial sensitivity testing was undertaken using the disc diffusion method following British Society for Antimicrobial Chemotherapy (BSAC) guidelines. Cefpodoxime resistance was used as a proxy for production of extendedspectrum beta-lactamase enzymes. Coagulase-negative Staphylococci, Bacillus spp., diptheroids and alpha-haemolytic Streptococci other than S. pneumoniae were considered as contaminants. Acute and convalescent blood samples and urine samples were stored at -80°C for further testing.
Testing for disseminated TB was only carried out in those who were known to be HIV infected or of unknown HIV status. One mycobacterial blood culture (BACTEC Myco/F Lytic, Becton Dickinson, USA) was inoculated with 5-7ml of blood and cultured at 37°C in the TB laboratory of the Malawi College of medicine. It was inspected daily for the first 14 days and then every two days thereafter for fluorescence with a handheld Wood's lamp. On the detection of fluorescence the contents were centrifuged and examined by Gram's and ZN stain to exclude contamination, and inoculated into mycobacterial growth indicator tubes (MGIT, Beckton Dickinson, USA) for up to 6 weeks. Subsequently cultured isolates were classified as M. tuberculosis or non-tuberculous Mycobacteria by microscopic cording and MPT-64 lateral flow assays (TAUNS Laboratories, Japan). After 8 weeks if no fluorescence was seen then centrifugation and microscopy was carried out; if no organisms were seen then this was classed as "no growth". 3 Batched urinary lipoarabinomannan testing using lateral flow assays (Alere determine TB, Abbot Laboratories, USA) was carried out on thawed frozen (-80°C) urine as per the manufacturer's instructions. Urine was thawed, briefly centrifuged to remove sediment and 60 microlitres of urine was applied to the sample pad and the result read after 25 minutes by comparing to the provided reference scale card. If a line was visible in the patient window of the test and darker than the lightest positive line on the reference card then it was considered positive. If a line was visible but lighter than the lightest positive line on the reference card, or if no line was visible, then it was considered negative. If no line was visible in the control window of the test then the test was considered invalid, and repeated. The results were read independently by two readers, who were unaware of the other reader's finding. A tiebreak read by a third reader who was unaware of the findings of the other two readers was undertaken in the event of disagreement.
Testing of sputum for tuberculosis using Xpert MTB/RIF and microscopy and culture of CSF were carried out when there was suspicion of pulmonary tuberculosis or meningitis, respectively, by the clinical team. Sputum Xpert MTB/RIF testing (Cepheid, United States) was carried out as per the manufacturer's instructions in the QECH hospital laboratories: sputum specimens were mixed with sample reagent and incubated at room temperature for 15 minutes. The liquefied specimen was then loaded into the Xpert MTB/RIF test cartridge for processing. CSF culture was carried out at the MLW laboratories: CSF was cultured on blood, chocolate and Sabouraud agar and then identification of any growth undertaken as for blood cultures above. In addition CSF was tested for cryptococcal antigen by lateral flow assay (Bruker Corporation, USA) as per the manufacturer's instructions.  Table 1). 122 serum samples were taken forward for testing; these were selected from all available acute sera by random sampling weighted 10:1 in favour of samples in which no diagnosis had been made by culture, urinary LAM testing, or sputum Xpert MTB/RIF. Samples were processed within a containment level 3 laboratory in preparation for sample extraction. Once processed, selected samples were then extracted on a MagNA Pure 96 instrument, stored in 2ml microtubes and frozen until tested. Samples were then processed on TaqMan Array cards within a containment level 2 laboratory using a defined working method. All positive hits were recorded on a master result sheet, including specific card internal controls.
Extraction controls were also processed and recorded. Table 1: TaqMan array card layout for the multipathogen panel. The microfluidic cards contain 384 individual wells arranged in 8 channels with 48 wells per channel. The design of the multipathogen array card was maximised to contain primer sets to detect 50 individual bacterial and viral pathogens. Channels 1, 3, 5 and 7 contain mainly primers against bacterial targets and channels 2, 4, 6 and 8 contain primers against viral targets. Each channel also contains RNAse P and MS2 internal controls. For several pathogens, more than one genetic target is included (indicated by #

Sample size
The study was powered to detect clinically relevant risk factors for death; we made the a priori assumption that a risk ratio of 2 or more is clinically relevant in this setting and patient population. We initially aimed to recruit 250 patients; with this number of patients and assuming a case fatality ratio of 50% (as seen in previous studies in Malawi) would have 80% power to detect risk factors for death with prevalence of 20-50% that confer a risk ratio (RR) of 1.5-3; therefore, a RR of 2 is likely to be detected. Resource considerations during the study resulted in a reduction of a feasible sample size to 225; repeating a power calculation with this number of participants suggested 80% power to detect risk factors conferring a RR of 2 with 25-50% prevalence, assuming 50% mortality.

Summary statistics
Continuous values were summarized by medians and interquartile ranges, categorical variables by proportions and exact binomial confidence intervals.
Differences between groups were assessed with bootstrapped differences in median or proportion with 95% confidence intervals constructed using the bias-corrected and accelerated bootstrap interval 2 . Where carried out, tests of difference in proportions between two groups use Fisher's exact test.
Estimating proportion of infections sensitive to ceftriaxone 7 To estimate the proportion of microbiologically infections that would be sensitive to ceftriaxone we used the actual antimicrobial sensitivity testing if that were available (for aerobic culture isolates); if not, we assumed that Streptococcus pneumoniae, Salmonella spp., Enterobacterales, Acinetobacter spp. were susceptible to ceftriaxone and Pseudomonas spp. and Enterococcus spp. were nonsusceptible. In addition, regardless of antimicrobial sensitivity testing result we assumed that organisms with an inducible chromosomal AmpC (Enterobacter, Serratia, Citrobacter, Aeromonas, Provadencia, Morganella spp.) were nonsusceptible to ceftriaxone. We assumed that Leptospira spp. were susceptible to ceftriaxone.

Estimating effects of treatments on outcome
We aimed to identify the effect of treatment on death by 28 days. We first hypothesized a causal structure (Supplementary Figure E1) of host variables (e.g. HIV, CD4 count, haemoglobin), infection (e.g. causative pathogen) and sepsis severity (e.g. lactate, blood pressure). We then used the R package dagitty 3 to identify that, to identify the causal effect of treatment variables on outcome, we need to condition on host, infection and severity variables, conditional on the hypothesized causal structure being correct. To account for collinearity of severity and host variables and to undertake dimensionality reduction on the large number of putative variables we used principal components analysis after log-transforming heavily skewed variables (Supplementary Table E1 (TB, invasive fungal disease, and malaria). We also included time-to-antibacterial therapy (defined as the time from ED registration to receipt of any antibacterial) as a continuous variable both as a linear variable and represented with restricted cubic splines with 3 knots at the centiles {0.1,0.5,0.9}; volume of intravenous fluid received was also modelled with the same spline function. Both variables were mean centred and scaled by their standard deviation before being included in the models.
The outputs from the models are presented as odds ratios (OR) with a point estimate (the posterior median) with 95% credible intervals (CrI), except for the nonlinear models which are presented as the marginal effect of the predictor on the probability of outcome with 95% credible intervals. Presence or absence of arboviral infections was not included in any mortality models as diagnostics testing was largely done only on convalescent samples.