Comparison of anti-malarial drug efficacy in the treatment of uncomplicated malaria in African children and adults using network meta-analysis

Published literature was searched in PubMed. The search included all potentially relevant published articles, characterized by random allocation to treatment and comparison of different artemisinin-based combinations in Africa, starting from the year of the introduction of ACT (2002–2003), up to June 2016. An electronic search of Medline was conducted using the following key words: (“malaria” AND ACT [All Fields]) AND (Randomized Controlled Trial [ptyp]) AND (“2002/01/01” [PDAT]: “2016/06/30” [PDAT]) AND “humans”. In addition, reference lists of reviews were also screened to include potential articles. Studies were eligible if they involved at least 2 ACT medicines and reported clinical efficacy as ACPR corrected by polymerase chain reaction (PCR). Cure, i.e. the term “adequate clinical and parasitological response”, was defined as undetectable parasitaemia with or without fever, without previously meeting the criteria of treatment failure on the last day of follow-up (usually day 28).

A total of 91 articles was identified and screened. Details are presented in Fig. 1. Several articles were excluded for the following reasons: studies that were either on tolerability or recurrence of parasitaemia or did not report data on ACPR [24, 25], or based on the mechanism of resistance to ACT [26], or comprised only one ACT-arm as in Ref. [27], which compared AL to azithromycin–chloroquine (AZCQ). Data comprising monotherapies and ACT were excluded. In addition, reference lists in 16 reviews were screened for recent studies [12, 13, 22, 28, 29]. At the end of the screening process, 76 articles were included in the original analysis; K = 13 treatments were involved. The quinine arm was removed from a 3-arm study [30], and two studies published in French were included [31, 32]. The list of the interventions was artesunate–amodiaquine (ASAQ), artesunate–sulfadoxine–pyrimethamine (ASSP), artesunate–sulfamethoxypyrazine–pyrimethamine (ASSMP), artemether–lumefantrine (AL), dihydroartemisinin–piperaquine (DHAP), dihydroartemisinin–piperaquine–trimethoprim (DHAPT), artesunate–chlorproguanil–dapsone (ASCD), ASMQ, artesunate–atovaquone–proguanil (ASATPG), artemisinin–naphtoquine (ASNAPH), artesunate–amodiaquine–chlorpheniramine (ASAQCPH), artesunate–pyronaridine (ASPY) and, the non-ACT combination amodiaquine–sulfadoxine–pyrimethamine (AQSP). The combination AQSP was included in the present analysis because this combination was employed during the transition period before the adoption of ACT in many African countries and its efficacy had been compared to that of ACT in randomized studies.

Patient populations included children and adults: 82% were children less than 15 years old among whom more than 80% were children under 5 years. Some published articles involved both children and adults or adults only. Trial characteristics are described in the Additional file 1. Most of the studies presented the outcome on day 28 only, while few followed the patient until day 42 or even day 63. Hence, only a few studies had longitudinal data on days 28, 42 and 63. Accordingly, the obtained information was pooled on the primary end-point day 28 using the observed number of ACPR. The percentage of ACPR was estimated using the number of enrolled population or the number followed as far as the intention-to-treat (ITT) analysis or a per protocol (PP) approach is concerned, i.e. when the ITT results were not reported, the percentage of ACPR obtained in a PP approach was used to estimate the number of patients with positive outcome.

A network of treatments was built after extracting information from the articles. Figure 2 shows the set of nodes represented by the name of ACT or AQSP, linked by lines. The thickness of the line is proportional to the number of randomized clinical trials that have been included. The lines joining AL-DHAP and AL-ASAQ are the thickest. The Figure has a complex structure and allows indirect comparison, e.g. ASMQ and ASSMP with ASAQ as the reference. Close loops like AQSP, ASAQ and ASMQ provide both direct and indirect evidence. The network also shows that AL was the most studied ACT. It could be used as the comparator treatment. However, ASAQ was the most commonly used ACT in clinical practice in Africa due to its cost, availability and tolerability. It has been shown in several randomized clinical trials that ASAQ is comparable to AL. Accordingly, AL was used as the entire network common comparator among trials as it has been evaluated against the highest number of treatment. Therefore, the network trial generated K – 1 = 12 contrasts representing the overall relative treatment difference to be estimated from the K (K − 1)/2 = 78 direct and indirect overall comparisons.

The numbers of multi-arms studies were as follows: 2-arm (64 studies), 3-arm (9 studies), and 4-arm (3 studies). The numbers of studies per-treatment arm were as follows: 12 AQSP, 63 AL, 42 ASAQ, 1 ASAQCPH, 1 ASATPG, 4 ASCD, 6 ASMQ, 2 ASNAPH, 2 ASPY, 4 ASSMP, 11 ASSP, 22 DHAP, and 1 DHAPT. A detailed summary regarding the number of studies per-treatment arm, randomized sample sizes, number evaluated for the outcome, and proportion of ACPR is given in Table 1 for each treatment. For example, AQSP was found in 12 articles but only 7 compared its efficacy to that of AL. All cure rates ranged from 64.6 to 98%.

Based on the information obtained from the data set, the drugs were numbered as follows except for AL, i.e. AL is treatment number 1 representing the common comparator drug, 2 = AQSP, 3 = ASAQ, 4 = ASAQCPH, 5 = ASATPG, 6 = ASCD, 7 = ASMQ, 8 = ASNAPH, 9 = ASPY, 10 = ASSMP, 11 = ASSP, 12 = DHAP, and 13 = DHAPT. For each study, data were summarized in terms of the number of treatment arms, list of treatments, number of PCR-corrected day 28 ACPR in each treatment arm, and sample size in each arm.

In order to simultaneously compare all treatments in a coherent manner, NMA was used to obtain a hierarchy of the competing interventions [15]. This method required assumptions, notations and presentation of the modelling approaches.

The application of NMA to the selected data assumed that the randomization process was preserved within each trial comparing the estimates of relative effect among treatments. Different assumptions were considered in NMA: (i) the hypothesis of homogeneity, i.e. no variation in treatment effect between trials within pairwise contrast; (ii) the consistency assumption establishing that within a closed loop, estimates for a particular pairwise contrast from head-to-head evidence is the same as what is estimated from the indirect evidence, and (iii) the inconsistency assumption, which means that estimates for a direct evidence are different from the indirect evidence.

For the following, the modelling approaches presented in a technical support document conceived by Dias et al. [33], was used. It is based on a random effect model with homogeneous variance, which allows assessing the validity of the consistency assumption and adjustment for multiple-arm trials. Notations described by Greco and Hong [34, 35] were followed to perform NMA for binary outcome using a Bayesian approach on aggregate data. The principal summary measure was the odds ratio (OR).

Let N = 76 and K = 13 denote the number of randomized clinical trials and the total number of treatments, respectively. Given that each study i included na _i number of treatment arms, i = 1,…, N, and the number of PCR-corrected ACPR r _i,k in every arm k (k = 1, …, na _i  ≤ K) of study i, then r _i,k follows a binomial distribution denoted ~Binom (p _i,k, n _i,k) , where n _i,k is the sample size of arm k in study i, and p _i,k is the probability of an event (success, here ACPR) of arm k in study i. For K and N, the number of existing comparisons (edges) between the treatments is equal to 109. This number was obtained by cumulating the number of pairwise comparisons among trials, i.e. the na _i (na _i  − 1)/2 comparisons. For instance, given a comparator within a trial, a 2-arm trial contributes one pairwise comparison, a 3-arm trial, three comparisons and a 4-arm trial, six possible comparisons. Given the K = 13 treatments, K (K − 1)/2, i.e. 78 treatment effects, are expected. The baseline (also called basic parameters) K – 1 = 12 has to be estimated. Let the following \(d_{12} , d_{13} , d_{14} , d_{35} , d_{16} , d_{17} , d_{18} , d_{19} , d_{1,10} , d_{1,13}\), \(d_{1,12} , d_{1,13}\) denote the baseline parameters, representing the overall treatment effect, with d ₃₅ referring to the direct effect between ASAQ and ASATPG. The probability of event occurrence p _ik is modelled on the logit scale as:where \(\mu_{i}\) is the “study” random effect that accounted for differences among trials. This is also the trial-specific baseline and represents the log-odds ratio (LnOR) of event in the overall control treatment, while \(\delta_{i,bik}\) is the trial-specific LnOR of event occurrence of the treatment k compared with the “placebo” treatment b _i , which can be considered here as a random variable that follows a normal distribution N (d _bk, \(\sigma_{i,bk}^{2}\) ), assuming that b _i  = b for all studies. Therefore, treatment indexed k = b = 1 corresponds to AL, and d _1,1  = d _k,k  = 0. In the context of NMA, it is assumed that the study-specific treatment effects are exchangeable such that \(\sigma_{i,bk}^{2}\) = τ ² , ∀ b,k; i.e. τ², the between studies variation, is assumed to be the same for all pairwise contrasts in all subsequent methods. \(\delta_{i,bik} \sim N\left( {d_{bk} ,\tau^{2} } \right)\) is seen as the common form of a random effect meta-analysis, especially in the frequentist methods. Here, the parameter \(\tau^{2}\) explains the variability among trials that could be caused by different drug formulations, the methods used, study designs, and the level of transmission that varies from one trial to another. In addition, the node \(\sigma_{i,bk}^{2}\) expresses the same variability within a study among a pairwise contrast.

To account for heterogeneity in the patient populations, a dummy variable called S _i was defined as 1 if the population was children less than 15 years and 0 in other patient populations. A random coefficient β_i for S _i was assumed. Hence, β_i followed a normal distribution with mean β₀ and variance ε².

The model is given by:where the node \(\delta_{i,bk}\) is a random effect with mean d _bk representing the treatment effect of b compared to k for all studies (k ≠ b) and a variance \(\sigma_{i,bk}^{2}\). This model provided the estimates of the K − 1 = 12 “basic parameters” representing all treatment relative to treatment b, by considering b = 1, 3. The parameter \(\sigma_{i,bk}^{2}\) can be expressed in terms of the study variation τ and the number of “basic parameters”, i.e. \(\sigma _{{i,bk}}^{2} = \tau ^{2} \times {{2\left( {K - 1} \right)} /K}\). For purpose of illustration, to derive the treatment difference between 2 = AQSP and 4 = ASAQCPH given AL as the control group, the consistency equation is given by d ₂₄  = d ₁₄  − d ₁₂ .

While heterogeneity is characterized as between-trial variation within treatment contrast, inconsistency is the variation between contrasts. Contrary to the consistency model, there is no common comparator across studies. Several methods have been studied to test the inconsistency assumption and are considered as an extension of Bucher’s method [36]. Instead of the repeat application of the method, Dias et al. [33] proposed to compare the consistency model with an inconsistency model. Hence, for a study comparing treatment k to treatment k ^’ , the model was defined as follows:

In this model, within study i, each node \(\delta_{{i,kk^{\prime}}}\) is a random variable with a mean treatment effect \(d_{{kk^{\prime}}}\) with a variance of \(\sigma_{i,bk}^{2}\). \(\delta_{{i,kk^{\prime}}}\) was treated as a separate parameter to be estimated. Accordingly, the model gives estimates of 12 overall treatment effects and also provides estimates of other 66 treatment effects. To test if consistency is reasonable, model 3 was compared to model 2 by checking statistical difference of their model fit.

Since there are 2-, 3- and 4-arm trials, the difference in treatment effects may arise from different sources and can be inconsistent [33]. Although only a small proportion of studies was 3-arm (n = 9) or 4-arm (n = 4) trials, with model 2, an adjustment for multiple arms trials was made. Indirect comparisons accounted for the correlation between any two treatment contrasts in a multi-arm trial. This correlation is equal to 0.5 given the hypothesis that the study variation is the same among all pairwise contrasts.

First, pairwise random effect meta-analyses were conducted. The method was implemented with the meta-analysis package of the software R (rmeta) downloadable at https://​cran.​r-project.​org/​. Treatments involved were the five WHO-recommended ACT medicines: ASAQ, ASSP, AL, DHAP and ASMQ. Direct estimates were extracted from head-to-head trials using the odds ratios and 95% confidence interval (CI). The heterogeneity of variance for each pairwise comparison was estimated using the \(I^{2}\) statistic.

Secondly, data were analysed by Bayesian method. Prior information was defined for all unknown parameters d _bk and \(\tau^{2}\). The following prior distributions were considered: \(\mu_{i} \sim N\left( {0, 0.0001} \right);\, d_{bk} \sim N(0, 0.0001)\). The inverse of the study variance was given a uniform distribution, i.e. \(s^{2} = \frac{1}{{\tau^{2} }}\sim Uniform(0,2)\); β0 ∼ N (0,0.0001), and 1/ε² also followed a uniform distribution dunif (0,2). One markov chain was run. After a burn-in of 100,000 iterations, posterior outputs were obtained from the last 50,000 iterations. Model fit statistics were obtained for each model, i.e. deviance information criterion (DIC), deviance (\(\hat{D}\)) which provides an idea of the model likelihood, and \(p_{D }\) the number of parameters in the estimation process. The model with the smallest DIC is considered the best. Regarding treatment ranking, the probabilities that each treatment is the best, second best, and third best on Day 28, among 13 treatments, were calculated according to van Valkenhoef and Kuiper [21]. Treatment ranks were based on posterior probabilities. In each MCMC run, every treatment was ranked according to its estimated magnitude. The proportion of MCMC cycles in which the treatment k ranks first yielded the probability that such specific treatment is the best among all treatments. Analyses were carried under the WinBUGS software. The WinBugs codes that were used for NMA are available in the Additional file 2.

Thirty-one articles (15,695 participants) compared the efficacies of ASAQ (taken as reference) and AL. The summary measure for these studies showed that the efficacy of ASAQ was comparable to that of AL (OR = 0.96; 95% CI [0.79; 1.17]; p value = 0.682; τ² = 0.138; I ²  = 54% [32.2%; 69.9%]). The efficacy of ASAQ (as reference) was compared to that of DHAP in six studies (4042 participants); ASAQ appeared to be less efficacious than DHAP but the result was not statistically significant (OR = 0.81; 95% CI [0.54; 1.22]; p value = 0.31; I ²  = 59.5% [0.5%; 83.5%]). Seven studies (2245 patients) compared ASAQ (reference) to ASSP. Results showed that both treatments were comparable (OR = 0.77; 95% CI [0.52; 1.14]; p value = 0.191; τ² = 0.055; I ²  = 26.7% [0%; 68.2%]). Eighteen articles (100,000 participants) compared AL (taken as reference) to DHAP. The results showed that AL was statistically less efficacious than DHAP (OR = 0.52; 95% CI = [0.36; 0.77], τ² = 0.518, p value = 0.0009; I ²  = 86.8% [80.6%; 91%]), but AL was as effective as ASMQ in five studies (2869 patients; OR = 0.90; 95% CI [0.56; 1.47], τ² = 0, I ²  = 0% [0%; 73.4%]). Three studies (1104 patients) compared AL and ASSP, and the differences were not statistically significant (OR = 1.14; 95% CI [0.40; 3.22], I ²  = 36.4% [0%; 78.7%). Only one study compared ASAQ to ASMQ. No study compared ASMQ to DHAP or ASSP. In addition, no study compared ASSP to DHAP.

Four trials (1584 participants) compared ASAQ (reference) to AQSP and showed that ASAQ was more efficacious than AQSP, but the results were not significant (OR = 1.51; 95% CI = [0.80; 2.87]; p-value = 0.20; τ² = 0.23, I ²  = 63.4% [0%; 87.7%]). AQSP (reference) was found less efficacious than AL in seven trials (2316 participants) but the result was not statistically significant (OR = 0.76; 95% CI [0.25; 2.30]; p value = 0.62; τ² = 1.89, I ²  = 93.7% [89.5%; 96.3%]) [37‐43]. The same conclusion was found in three trials comparing AQSP (ref) to DHAP (2518 patients; OR = 0.54; 95% CI [0.16; 2.03]; τ² = 1.30, I ²  = 96.6% [93%; 98.4%]) [42, 44, 45]. Forest plots of some of these results are illustrated in Additional file 3.

Table 2 displays the overall treatment effects (given AL as the common comparator) estimated from two models using the Bayesian approach and their 95% credible interval (CrI). All models yielded similar results. The difference in DIC of models 3 and 2 was 1025.04–1024.07 = 0.97, suggesting little evidence for inconsistency. Hence, model 2 gave the best adjustment to the data set. A significant difference was found between DHAP and AL. For example, with model 2, DHAP was 1.92-fold more efficacious than AL (OR = 1.92; 95% CrI = 1.30–2.82; 63 vs 22 studies per-treatment arm; 19,163 participants). New drug combinations like ASSMP, ASATPG, ASPY, and ASAQCPH appeared more efficacious than AL, but the results were not statistically significant as indicated by the confidence intervals. For model 3, the other 78 − 12 = 66 comparisons can be found in Additional file 3.

Indirect estimates were calculated from each model assuming the consistency equation (Table 3). The results showed the superiority of DHAP compared to ASCD (OR = 3.25, 95% CI 1.46–7.25), ASAQ (OR = 1.70; 95% CI 1.10–2.64), and AQSP (OR = 2.20; 95% CI 1.21–3.96). Indirect estimate of DHAP vs ASMQ yielded OR = 1.59 with 95% CI 0.62–4.02 showing no significant difference between the two drugs. The efficacy of DHAP was also not statistically different from that of ASPY (OR = 1.36; 95% CI 0.38–4.87). The combination ASATPG was 3.40-fold more efficacious than AL, albeit not statistically significant (95% CI 0.60–18.56).

Model 2 was selected for calculating the rank probabilities. Table 4 displays the rank for each treatment and the corresponding rank probability. Data are shown for rank 1, rank 2, rank 3 and for the worst treatment (rank 13). ASATPG combination had the highest probability (0.53) at rank 1, followed by ASNAPH (0.11) and ASAQCPH (0.11). At rank 2, DHAP was the drug with the highest probability (0.25). At rank 3, the best was also DHAP (0.31). For rank 13, the worst treatment was ASCD (0.48).

To provide an answer to the best treatment among WHO-recommended ACT medicines and AQSP, a second analysis was carried out using model 2, with AL as the overall control group. Treatments were numbered as follows: AL = 1, AQSP = 2, ASAQ = 3, ASMQ = 4, ASSP = 5, DHAP = 6, 7 = ASAQCPH, 8 = ASATPG, 9 = ASCD, 10 = ASNAPH, 11 = ASPY, 12 = ASSMP, and 13 = DHAPT, allowing the WHO-recommended ACT medicines to be the first 6 treatments. Ranking was only up to 6. Table 5 displays the results. DHAP was more efficacious than AL (OR = 2.09; 95% CI [1.54–2.83]). At ranks 1, 2, and 3, DHAP emerged at the first treatment rank.