Serious adverse events following treatment of visceral leishmaniasis: A systematic review and meta-analysis

Background Despite a historical association with poor tolerability, a comprehensive review on safety of antileishmanial chemotherapies is lacking. We carried out an update of a previous systematic review of all published clinical trials in visceral leishmaniasis (VL) from 1980 to 2019 to document any reported serious adverse events (SAEs). Methods For this updated systematic review, we searched the following databases from 1st Jan 2016 through 2nd of May 2019: PUBMED, Embase, Scopus, Web of Science, Cochrane, clinicaltrials.gov, WHO ICTRP, and the Global Index Medicus. We included randomised and non-randomised interventional studies aimed at assessing therapeutic efficacy and extracted the number of SAEs reported within the first 30 days of treatment initiation. The incidence rate of death (IRD) from individual treatment arms were combined in a meta-analysis using random effects Poisson regression. Results We identified 157 published studies enrolling 35,376 patients in 347 treatment arms. Pentavalent antimony was administered in 74 (21.3%), multiple-dose liposomal amphotericin B (L-AmB) in 52 (15.0%), amphotericin b deoxycholate in 51 (14.7%), miltefosine in 33 (9.5%), amphotericin b fat/lipid/colloid/cholesterol in 31 (8.9%), and single-dose L-AmB in 17 (4.9%) arms. There was a total of 804 SAEs reported of which 793 (including 428 deaths) were extracted at study arm level (11 SAEs were reported at study level only). During the first 30 days, there were 285 (66.6%) deaths with the overall IRD estimated at 0.068 [95% confidence interval (CI): 0.041–0.114; I2 = 81.4%; 95% prediction interval (PI): 0.001–2.779] per 1,000 person-days at risk; the rate was 0.628 [95% CI: 0.368–1.021; I2 = 82.5%] in Eastern Africa, and 0.041 [95% CI: 0.021–0.081; I2 = 68.1%] in the Indian Subcontinent. In 21 study arms which clearly indicated allowing the inclusion of patients with HIV co-infections the IRD was 0.575 [95% CI: 0.244–1.355; I2 = 91.9%] compared to 0.043 [95% CI: 0.020–0.090; I2 = 62.5%] in 160 arms which excluded HIV co-infections. Conclusion Mortality within the first 30 days of VL treatment initiation was a rarely reported event in clinical trials with an overall estimated rate of 0.068 deaths per 1,000 person-days at risk, though it varied across regions and patient populations. These estimates may serve as a benchmark for future trials against which mortality data from prospective and pharmacovigilance studies can be compared. The methodological limitations exposed by our review support the need to assemble individual patient data (IPD) to conduct robust IPD meta-analyses and generate stronger evidence from existing trials to support treatment guidelines and guide future research.


Introduction
Visceral Leishmaniasis (VL) is a vector-borne disease caused by protozoan parasites of the genus Leishmania. The disease is endemic in parts of South Asia, East Africa, South America and the Mediterranean region with an estimated 50,000 to 90,000 cases occurring annually [1]. VL is mainly caused by L. donovani in East Africa and the Indian Subcontinent (ISC), and by L. infantum in the Mediterranean region and South America [2]. VL continues to exert substantial burden and is a leading cause of morbidity in the regions where the disease is endemic [3]. Untreated symptomatic forms of VL are almost always fatal, with the mortality dropping to approximately 5% among those who are treated [4]. and a biomedical resource grant from Wellcome Trust to PJG (ref:208378/Z/17/Z). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: We have read the journal's policy and the authors of this manuscript have the following competing interests: PLO was previously a staff member of the World Health Organization. CH is a staff member of the World Health Organization. All other authors declare that they have no competing interests. The authors alone are responsible for the views expressed in this publication and it does not necessarily represent the decisions, policy, or views of their respective organisations.
The safety of antileishmanials remains a distinct cause of concern despite their success in achieving a clinical cure, especially when deployed at scale in routine control programmes. The poor tolerability of VL therapeutic options was established as early as the first decades of the 20 th century. Empiric treatment with tartar emetic, a trivalent antimony, produced toxicity that outweighed their therapeutic benefit to the patients in several cases [5]. Dose-dependent toxicities including prolonged QT interval, cardiac arrhythmia [6,7] and Torsade de Pointes [8] have been reported following pentavalent antimony (PA), the successor of trivalent antimony. Despite these known toxicities, absence of alternatives meant that PA remained the mainstay treatment against VL for much of the 20 th century. Several therapies were tested but safety continued to be a major hurdle in drug development: pentamidine which was introduced in the 1950s was associated with triggering insulin-dependent diabetes mellitus [9]; sitamaquine was associated with glomerulonephritis, which caused its development to be terminated [10,11]; unacceptable toxicity precluded the use of amphotericin B deoxycholate among patients with renal impairment and confined its usage mainly to teaching and district hospitals in India [12]. The discovery that encapsulation of deoxycholate into a liposomal membrane (liposomal amphotericin B) allows a lower drug exposure in the plasma leading to improved safety heralded a new era in the treatment of VL [13].
The liposomal form of amphotericin B has since been explored in several trials in Asia, Africa and Latin America and is well tolerated [4]. It has been tested in three different formulations: colloidal dispersion amphotericin B, lipid complex amphotericin B, and liposomal amphotericin B (L-AmB). The latter is currently deployed either as a single-dose regimen (10 mg/kg) [14] or as multiple doses administered over a treatment period often lasting 3-10 days (with a total dose up to 40 mg/kg) [2,15]. The practical advantages offered by the single-dose L-AmB regimen (10 mg/kg), coupled with high efficacy and acceptable safety profile in clinical trial settings, has led it to become the regimen of choice as the frontline drug in the ongoing Kala-Azar elimination programme (KAEP) in Bangladesh, India, and Nepal [16]. The KAEP requires that patients are followed-up for at least six months post-treatment to gauge events like relapse, death or any other serious adverse events (SAEs) (serious adverse events are defined as any undesirable experience associated with the study drug that leads to the following patient outcomes: life-threatening, prolonged hospitalisation, death, disability/permanent damage, or congenital/birth defect), and up to 3 years of follow-up is recommended to capture the incidence of post-kala-azar dermal leishmaniasis (PKDL) [17]. In practice, due to resource constraints, only few cases are followed-up. In addition to the study follow-up, the intrinsic inter-and intra-regional variation in drug efficacy [18], co-infection with HIV [4], and regional variation in age-sex distribution [19,20] might impact the safety and tolerability profile of the drugs. For example, among patients with HIV co-infections, Liposomal amphotericin B either as monotherapy or in a combination with miltefosine generally have been associated with favourable outcomes [2,15,21,22]. While the safety of antileishmanial is wellstudied in clinical settings, only a handful of studies have characterised the tolerability profile of the drug when used in non-clinical trial settings [23,24]. If the expected event rates after drug administration in a controlled clinical trial setting are known, this might provide a baseline measurement for a more targeted surveillance in community settings and may serve as a benchmark for future trials against which mortality data from prospective and pharmacovigilance studies can be compared.
This systematic review of published clinical trials was conducted with an overarching aim of documenting reported serious adverse events (SAEs) and quantifying the observed mortality rate following drug administration in patients with visceral leishmaniasis.

Eligibility criteria
A systematic review of the existing scientific literature to identify clinical trials of therapies for VL was conducted in accordance with the Preferred Reporting Items for Systematic-Reviews and Meta-Analyses (PRISMA) guidelines (S1 Text) [25]. A pre-determined inclusion and exclusion criteria, described elsewhere, were used for study screening [26]. Briefly, all clinical trials in humans assessing the efficacy and safety outcomes of antileishmanial drugs as intervention were eligible for inclusion in this review. The review was not limited by comparators or language given the desire to capture all trials evaluating any treatment modality. Studies on PKDL, cutaneous leishmaniasis, canine visceral leishmaniasis, and those evaluating vaccines or prophylactic usage of drugs, vector control, net distributions, prevalence estimation and diagnostic tests were excluded.

Information sources and search strategy
The identification of studies included in this study is the result of two separate literature searches. This review builds on a previous systematic review which had identified clinical trials registered or published before January 2016 [26]. The current search was designed to capture any VL clinical trial records published or registered between 1 st January 2016 and 2 nd of May 2019 (search was run on 2 nd of May 2019) in the following databases: PubMed, Ovid Embase, Scopus, Web of Science, Cochrane, clinicaltrials.gov, WHO ICTRP, and the Global Index Medicus. The update search was run by an experienced librarian at the Bodleian Health Care Libraries, the University of Oxford. The systematic review considered both randomised and non-randomised interventional studies including single-armed trials for the assessment of the serious adverse events (SAEs) [27]. Details of the search strategy adopted for each of the databases are provided in supplemental file (S2 Text).

Study selection and data extraction
Two reviewers (MB, SR) screened the studies identified in the search in a blinded fashion using Covidence software for systematic reviews [28]. Any disagreement was resolved by contacting a third reviewer (BJM). Variable and data dictionaries with extraction rules were developed for the construction of a REDCap electronic data capture tool hosted at the University of Oxford [29].
Data on the following aspects were extracted from studies meeting inclusion criteria: study design, location, follow-up duration, the number of participants enrolled, information on study arms, and drug posology. Data on the following safety parameters were extracted: reports of hypersensitivity/allergic reaction to treatment regimen at dose testing, reports of SAEs and death during treatment or study follow-up.
The REDCap database was piloted with an initial extraction of 10 studies by two authors (AH and SSP) with the aim to simplify and standardise extraction rules. Two authors (RN and SSP) independently validated/updated the variables already extracted in the previous review [26] and extracted data from all the articles included in this review (each author extracted data from half of the studies). Any discordances that developed in the course of the data extraction were resolved through discussions for clarity and consistencies of the data being extracted among the two reviewers. In case of discrepancies, a resolution was reached with further discussions with two authors (PD, BJM). A detailed list of variables and metadata is provided in supplemental file (S1 Data).

Statistical analyses and synthesis of results
Monitoring of adverse events in included studies. If there were any reports of AE (including death) during the treatment or the study follow-up; then the study was defined to have adverse events monitoring system (AEMS) in place. The standards used for coding/grading the AEs were extracted (when reported). If neither AEs were reported nor the standards were mentioned, the study was assigned "Unclear" AEMS status.
Hypersensitivity to the study drug. Data were extracted on dose-testing prior to the administration of amphotericin B formulations. The rate of allergic reactions leading to treatment discontinuation after dose testing was calculated using a random effects meta-analysis of single proportion using logistic transformation.
Serious adverse events (SAEs). Data on SAEs were extracted as reported in the original publications. Due to difficulties in accurately extracting the unique number of patients who experienced SAEs at any time during the study treatment and follow-up period, the analysis of SAEs was limited to reporting descriptive statistics.
Estimation of rate of death. When the time of death was not explicitly stated, description of death "during the therapy or treatment period" was considered as death occurring within the first month (� 30 days) of treatment initiation, and deaths after discharge or during follow-up were considered as occurring after day 30. Similarly, in studies not reporting any information on death, it was assumed that no deaths had occurred. The incidence rate of death (IRD) within the first 30 days of treatment initiation from individual treatment arms were combined using random effects Poisson regression and were expressed as the number of deaths per 1,000 person-days at risk [30]. Heterogeneity was assessed using I 2 statistics, which quantifies the proportion of total variability that is due to between-study differences [31]. The pooled estimates were presented together with the associated 95% confidence intervals (95% CIs). In addition, 95% prediction intervals (95% PIs) were also presented, which provides a range of values in which a future estimate of 30-day mortality rate in a set of VL patients will fall with probability of 0.95 [32]. The pooled incidence rates were adjusted using Copas selection model to account for potential publication biases [33,34]. Compared to other methods for testing publication bias (e.g., asymmetry of funnel plot), the Copas selection model explicitly models the publication bias by estimating the probability that a study is selected for publication conditional to the precision of the study result [33,34]. Mean difference in incidence rates were presented between sub-groups of interest and associated 95% confidence interval for the mean difference was constructed using delta rule (S3 Text). We further stratified the metaanalysis by whether the study allowed for inclusion of patients with HIV co-infections and geographical region.
Assessment of risk of bias. The assessment of the risk of bias in studies included in this review was carried out using the Cochrane Risk of Bias (ROB) tool for randomised controlled trials. Two reviewers independently assessed the risk of bias in randomised studies based on their design and conduct. Risk of bias in non-randomised studies were assessed using ROBINS-I tool (with modifications adopted for the current review) [35].
Sensitivity analyses. Sensitivity analysis was carried out by considering all deaths with unclear time of occurrence as occurring within the first 30 days of treatment initiation (worst case analysis). Estimated rates were also estimated separately for randomised and other than randomised studies (single-armed trials, non-randomised studies, alternative or consecutive treatment allocation or randomisation status not specified), and by the adjudicated risk of bias status on different domains.
Software. All statistical analyses were carried out using R software [36]. Meta-analysis of incidence rates was carried out using the metarate function in meta package and sub-group differences were explored using the test of moderator effects. Estimates adjusted for selection biases (using Copas selection model) were derived using the metasens package.

Study selection and characteristics
There were 2,271 records identified from the updated literature searches for the period 1 st January 2016 to 2 nd of May 2019, of which 1,330 were unique. Of the 1,330 unique records, 1,285 were excluded at the level of title and abstract screening. Full texts of the remaining 45 records were assessed, of which only 19 met the eligibility criteria. All 162 studies identified in the previous systematic review (period 1980 to 2015) [26], met the eligibility criteria and were merged with the 19 newly identified studies. The merger identified 173 (156 published; 17 ongoing) unique studies after the removal of 8 duplicates. One ongoing study was published and identified through personal communication leading to a total of 157 studies published from 1980 through to 2019 included in this review for data extraction (Fig 1).

Adverse events monitoring system
An adverse event monitoring system (AEMS) was reported to be in place for 149 (94.9%) studies whereas it was not clear for the remaining 8 (5.1%) studies. Of the 149 studies that reported AEs were systematically captured, the severity of the AEs were graded using Common Toxicity Criteria (CTC) of the National Cancer Institute in 34 (22.8%), WHO toxicity scale in 5 (3.4%), Division of AIDS Table for grading severity in 2 (1.3%), and the grading standards adopted was not described or unclear in the remaining 108 (72.5%) studies (See S1 Data).

Dose testing in amphotericin B formulation arms
History of hypersensitivity to the study drug was an exclusion criteria in 98 (28.2%) treatment arms of which 57 arms (58.2%, 57/98) tested amphotericin B (any formulation or in combination with other drugs). Out of 165 treatment arms (n = 14,195 patients) with amphotericin B  (any formulation or combination), dose testing prior to initiation of treatment regimen was reported in 38 (23.0%) arms (n = 4,594 patients). Two patients developed allergic reactions during a test infusion with 1 mg of L-AmB and were discontinued from the study. Using a random effects meta-analysis of single proportion using logistic transformation, the rate of allergic reactions leading to treatment discontinuation after dose testing was estimated at 0.4 cases [95% CI: 0.1-1.7] per 1,000 tested participants (I 2 = 0.0%; 38 arms, 4,594 patients).

Serious adverse events (SAEs)
A total of 804 SAEs could be extracted from the studies included in the analysis. In two studies, 11 SAEs were reported at the study level and could not be assigned to specific treatment arms leaving 793 SAEs, which could be identified at the treatment arm level.

Death reporting
There was a total of 439 deaths of which 428 could be identified to the specific treatment arm under study. Of the latter, 285 (66.6%) deaths were clearly described as occurring during the first 30 days of enrolment, 70 (16.4%) deaths occurred after completion of the treatment, i.e. between day 31 and the end of the study follow-up. In the remaining 73 (17.1%), the time of the occurrence of deaths could not be clearly ascertained. There were 60 (14.0%) deaths which were described as related to or possibly/probably/might be related to the drug administered ( Table 2). Further description of time and causes of deaths are reported in supplemental file (S1 Data).  (Table 3).

Estimate of the IRD within 30-day of treatment initiation stratified by enrolment of patients of HIV co-infections
There was a significant sub-group difference in the IRD when accounting for studies allowed inclusion of patients with HIV co-infection in a meta-regression (I 2 = 75.4%; P-value<0.001).  After stratifying by region, the IRDs for multiple doses of L-AmB, pentavalent antimony, and miltefosine were higher in Eastern Africa than in the Indian Subcontinent (Table 3).

Adjustment for potential publication biases
The incidence rates of deaths were re-estimated to account for potential publication biases; the adjusted rates were 2.6 to 36-fold larger compared to the unadjusted rates (Table 3 and Fig 2).

Sensitivity analyses and risk of bias assessment
In a worst-case scenario analysis, that considered all the deaths with unclear time as occurring within the first 30 days of treatment initiation, the overall estimated incidence rate of death was 0.069 [95% CI: 0.041-0.116; I 2 = 82.1%] per 1,000 person-days (S2 Table). The rates were estimated separately for randomised and non-randomised studies; the estimated rate was 0.046 [95% CI: 0.020-0.103] for the randomised studies and 0.129 [95% CI: 0.067-0.246] for non-randomised studies (P-value: 0.0011 for test of sub-group differences) (S3 Table). The IRD estimates were higher in studies considered to be at high risk of bias due to incomplete outcome reporting for both randomised and non-randomised studies (S4-S7 Tables). Similarly, risk of bias was also assessed in an eligible study that was initially excluded due to unavailability of full text [38]. The study tested miltefosine regimen in 160 children and reported 2 deaths within a week of treatment initiation. The IRD within 30 days was estimated at 0.416 [95% CI: 0.056-1.505]; the point estimate for IRD was larger but the 95% CI overlapped with the estimates for miltefosine for the Indian Subcontinent (Table 3).

Discussion
From this systematic review of all published antileishmanial efficacy clinical trials conducted in the last 4 decades, we report several findings directly relevant to researchers, field workers and policymakers regarding different aspects of drug safety. Hypersensitivity to the study drug is one of the major safety concerns in antileishmanial chemotherapies. Case reports of anaphylactic shock following amphotericin B has been welldocumented [39]. To obviate this, manufacturers recommend that patients are given a small test dose of L-AmB before the commencement of the full treatment course and hypersensitivity to L-AmB is often an exclusion criterion in clinical trials [40]. From 165 treatment arms with amphotericin B formulations included in this review, the estimated rate of discontinuations from the study after administration of a test dose was 0.4 [95% CI: 0.1-1.7] per 1,000 tested participants. This might, however, underestimate the true rate, since a history of drug hypersensitivity was an exclusion criteria in 57 arms (34.5%) that tested amphotericin B formulations, and reporting of dose testing information was available only in 38 (23%) study arms with this regimen.
There was a total of 793 SAEs that were reported at the study arm level; the majority of the descriptions regarding the cause were non-specific and the time of occurrence was missing in just over half, which precluded meta-analysis. As with the SAEs, the reporting of fatal outcomes was also inconsistent with the time of occurrence of death not documented in nearly a fifth of all reported deaths. Overall, mortality was an extremely rare event with the rate estimated at 0.068 [95% CI: 0.041-0.114; 95% PI: 0.001-2.779] per 1,000 person-days at risk with substantial statistical heterogeneity in results (I 2 = 81.4%).
Some of the heterogeneity was explained by geographical region as the overall estimated rate was much higher in Eastern Africa (0.628 [95% CI: 0.368-1.021]) than in the Indian Subcontinent (0.041 [95% CI: 0.021-0.081]) ( Table 3). For a given drug regimen, the estimates of the mortality rate across the regions were comparable for all the drug regimens with the exception of pentavalent antimony, multiple dose L-AmB regimen, and miltefosine, for which the estimated incidence rates were much higher in Eastern Africa (EA) than in the Indian Subcontinent (ISC). The larger rates observed in EA was partly explained by the clinical differences in patients enrolled as the studies from EA were more likely to enrol patients with HIV co-infectionns compared to studies from the ISC (16% in EA compared to 0.5% in the ISC); and the estimated mortality rate was greater than 10-fold higher in study arms which indicated patients with HIV co-infections were eligible for enrolment. This is consistent with published literature that the case fatality rate is much higher in HIV co-infected individuals compared to VL infections alone [41]. In patients with HIV infections, VL manifests as an opportunistic infection with disease severity exacerbated due to the depleted immune response, therefore with higher risk of poor treatment outcome and death. There is also an increased complexity in terms of case management as these patients are at an increased risk of drug toxicity due to interactions between antiretroviral and antileishmanial drugs, though these factors have not been extensively studied [42].
This review has several limitations. First, information regarding the time and causes of death and the relationship to the study drugs were found to be poorly described in the existing literature. This inconsistency in reporting made it difficult to interpret the data and prevented a direct comparison of the incidence rate of deaths between different regimens. We therefore carried out sensitivity analyses to explore the worst-case scenario where all the deaths with an unclear time of occurrence were considered as occurring within the treatment period; the estimates of IRD obtained from the sensitivity analyses were similar to the results reported in Table 3 (See S5 and S7 Tables). Like with mortality data, extraction of data on SAEs was difficult because most of the description were either not specific or the time when SAEs occurred was not available in just over a third of them. In particular, the underlying denominators were not clear as some patients could have experienced multiple SAEs. Second, despite there being only a few classes of antileishmanial drugs, there were many different formulations and dosages of amphotericin B or antimonies administered across the included studies. The multiplicity of drug regimens posed analytical challenges for evidence synthesis when using aggregated information from publications. Third, mortality is a rare event and several trials had no reported deaths, leading to a statistical problem of structural zeros. How best to handle such structural zeros in meta-analysis has remained a contentious issue in the statistical literature [30,[43][44][45][46]. The results presented in this review are derived from a random effects meta-analysis using Poisson regression model, following the recommendations of Spittal et al. [30]. Fourth, clinical aspects of the disease such as severity of illness and unresponsiveness of the parasites were not considered as data on these parameters could not be reliably extracted. Similarly, clinical trials by nature cannot detectsome SAEs like birth/congenital defects which can only be captured through prospective cohort studies/pregnancy registries. Finally, the Copas selection model indicated existence of publication bias as the adjusted estimates were larger than the unadjusted rates (Table 3 and Fig 2). The estimated rates were also affected by the heterogeneity in study design and by the risk of bias status in different domains (S5 and S7 Tables). The estimated rates should therefore need a cautious interpretation with these limitations taken into considerations.
Despite these limitations, this review reports several important findings which can help field workers in the assessment of the safety of antileishmanial therapies and guide further research with the data presented in this review serving as a baseline rate. This review has also found that reporting of safety outcomes in VL trials remains inconsistent, a finding in agreement to a previously published research [47]. In just over two-third of the studies included, the standards adopted for the gradation of the adverse events and serious adverse events were not mentioned. This suggests a need for clear and transparent reporting of the methodology used in safety assessment and warrants an adoption of recognised standards for reporting of safety outcomes such as Medical Dictionary for Regulatory Activities (MedDRA) or the Common Terminology Criteria for Adverse Events [48,49].
A carefully planned individual participant data (IPD) meta-analysis will ameliorate some of the limitations identified in this aggregated data meta-analysis and will allow the generation of a more robust estimate by amalgamating the current evidence base to the highest standard. The efforts of assembling and standardising IPD are ongoing under a data platform championed by the Infectious Diseases Data Observatory (see https://www.iddo.org/researchthemes/visceral-leishmaniasis). This is of utmost importance as the current research and development drug pipeline is limited, and improvement of safety reporting will also benefit future surveillance for new drugs that are under development [18]. Another factor to consider is the delivery of treatment at the "point of care" in national programmes which are often ill equipped in rural areas and lack trained staff to administer antileishmanial drugs. This real-life situation is much different than carefully planned clinical trials with mechanisms to monitor and address the incidence of adverse events. A better understanding of the safety and tolerability profile of existing treatment regimens is therefore critical for the assessment of therapeutic risk-benefit balance to facilitate judicious use of treatment. This can, in turn, prolong the shelflife of the current therapeutic arsenal at the frontline of elimination programmes and aid in combating drug resistance.
In conclusion, the overall expected incidence rate of death in the first 30-days after treatment initiation was a rarely reported event in clinical trials with the rate estimated at 0.068 per 1,000 person-days at risk [95% prediction interval: 0.001-2.779]. The estimated rates were higher in studies that allowed enrolment of patients with HIV co-infections and in studies from Eastern Africa compared to the Indian Subcontinent. The analysis of safety data was significantly inhibited by poor reporting and warrants a better and standardised reporting.