Conference Agenda

Background

Often when undertaking meta-analyses of time-to-event (TTE) outcomes, especially in Health Technology Assessment (HTA), a hazard ratio scale is used. However, issues arise when there is evidence of non-proportional hazards in some of the trials/studies included. When Individual Patient Data (IPD) are available or have been re-created from published Kaplan-Meier curves a number of methods have been advocated, including; flexible parametric models, piecewise exponential models, fractional polynomial models, and Restricted Mean Survival Time (RMST) models. However, their use has been limited by either their complexity and/or the ease with which their results can be incorporated into an economic decision model in order to assess cost-effectiveness.

An alternative approach is to assume a treatment-log(time) interaction within a Cox proportional hazards model in each trial/study, thus allowing the log HR to vary linearly with respect to log(time), and to then undertake a bivariate meta-analysis of the resulting treatment and interaction coefficients, so that an overall time-varying HR can be obtained.

Methods

A treatment-log(time) approach was applied to an IPD meta-analysis of 20 trials, involving 4,069 patients, of chemotherapy compared to Standard of Care (SoC) for advanced recurrent gastric cancer undertaken by the Global Advanced/Adjuvant Stomach Tumor Research International Collaboration (GASTRIC) Group, and in which Progression-free Survival (PFS) was an outcome with follow-up up to 6.8 years (2500 days). This approach was compared with a standard random effects meta-analysis of the trial-specific log HRs.

Results

Of the 20 trials in the meta-analysis 5 displayed evidence of non-proportional hazards for PFS. A standard random effects meta-analysis of HRs (undertaken on a log scale) yielded a pooled HR of 0.78 (95% CI: 0.71 to 0.86). Undertaking a bivariate random effects meta-analysis of the treatment and treatment-log(time) trial-specific coefficients produced a pooled interaction effect of +0.10 (95% +0.002 to +0.19) P=0.04 on a log hazard scale. The resulting HRs estimated at 250, 500, 750 and 1000 days were 0.85 (95% CI: 0.78 to 0.93), 0.91 (95% CI: 0.80 to 1.04), 0.95 (95% CI: 0.81 to 1.11) and 0.97 (95% CI: 0.81 to 1.17) respectively.

Conclusion

A treatment-log(time) interaction approach to the meta-analysis of TTE outcomes when the proportional hazards assumption appears not to hold for at least some of the studies included produces a simple and intuitive solution which can be readily incorporated into an economic decision model. Further extension to both a network meta-analysis setting and a fully Bayesian one-stage model is also possible.

Background: There is a growing demand to perform network meta-analysis (NMA) in health technology assessment (HTA) submissions, especially to meet the new EU-HTA requirements. Time-to-event clinical endpoints are commonly used in many disease areas. In NMAs with multiple studies, there is a high chance that proportional hazards (PH) assumption is not met for at least one study. In such situations, flexible methods for survival modelling are more suitable. Fractional polynomials (FPs) and Splines are commonly used flexible parametric models when PH assumption fails. FPs have a simpler mathematical form and require fewer parameters but may not capture all possible non-linear relationships. In contrast, splines are more flexible and can fit to more complex patterns, although they risk overfitting due to their high flexibility and consequently may not be suitable for extrapolation. However, recent developments in spline models have addressed the risk of overfitting. In this study, we assess the performance of FP and spline models.

Methods: We simulate time-to-event data for multiple studies within a NMA network following different underlying distributions including (a) simple parametric distributions (b) complex distributions to account for delayed response to treatment and the existence of long-term survivors. We fit various FP and spline models to this simulated network to compare their predictive and extrapolation performance for new data simulated from the same underlying data generating process.

Outlook: This study provides insights into the relative performance of FP and spline models in NMA and may help one to choose the appropriate flexible parametric modelling approach for time-to-event data.

Meta-analysis summarises the results of multiple independent clinical trials. For survival outcomes, both fixed effect and random effect meta-analyses are traditionally carried out using aggregated data obtained from published effect estimates. However, relevant data may not have been published for specific analyses, such as subgroups of interest, or administrative censoring may have occurred at different time points. In such cases, individual participant data (IPD) meta-analyses may be performed. The aim of this work is to provide a practical guide for the conduct of IPD fixed effect and random effect meta-analyses for survival outcomes in the presence of competing events. For the random effect analysis, we model optimal correlated frailties from the gamma distribution representing unobserved covariates at the cluster level, thus allowing for correlations between the different competing events within the studies. This presentation builds on previous work by Meddis et al. (2020 Biom J) and Rueten-Budde et al. (2019 Stat Med). It provides competing risks estimators for both cause-specific and subdistribution hazard ratios. For illustration, the IPD data from Meddis et al. are re-analysed. Specifically, data from 23 randomised controlled trials with a total of 4552 patients suffering from nasopharyngeal carcinoma and two competing events are considered. In summary, fixed effect and random effect meta-analyses can be performed with ease on both aggregated and IPD data.

Background: Interrupting human-to-mosquito transmission is important for malaria elimination strategies as it can reduce infection burden in communities and slow the spread of drug resistance. Antimalarial medications differ in their efficacy in clearing the transmission stages of Plasmodium falciparum (gametocytes) and in preventing mosquito infection. Here we present a combined analysis of six trials conducted at the same study site with highly consistent methodologies that allows for a direct comparison of the gametocytocidal and transmission-blocking activities of fifteen different antimalarial regimens or dosing schedules.

Methods and findings: Between January 2013 and January 2023, six clinical trials with transmission endpoints were conducted at the Clinical Research Centre of the Malaria Research and Training Centre of the University of Bamako in Mali. These trials tested Artemisinin-Combination Therapies (ACTs), non-ACT regimens and combinations with 8-aminoquinolines. Participants were males and non-pregnant females, between 5-50 years of age, who presented with P. falciparum mono-infection and gametocyte carriage by microscopy. Blood samples were taken before and after treatment for thick film microscopy, infectivity assessments by mosquito feeding assays and molecular quantification of gametocytes. A network meta-analysis (NMA) was performed to combine direct and indirect effects of treatment arms across studies. This analysis quantified changes in mosquito infection rates and gametocyte densities within treatment arms at 2 days, 7 days, and 14 days post-regimen relative to baseline (day 0). These quantified relative changes within arms were also compared between arms. In a pooled analysis of 422 participants, we observed substantial differences between antimalarials in gametocytocidal and transmission-blocking activities, with artemether-lumefantrine (AL) being significantly more potent at reducing mosquito infection rates within 48 hours than dihydroartemisinin-piperaquine (DHA-PPQ), artesunate-amodiaquine (AS-AQ), sulfadoxine-pyrimethamine plus amodiaquine (SP-AQ) and pyronaridine-artesunate (PY-AS) (p<0.0001). The addition of single low dose primaquine (SLD PQ) accelerated gametocyte clearance and led to a significantly greater reduction in mosquito infection rate within 48-hours of treatment for each ACT, while an SLD of the 8-aminoaquinoline tafenoquine (TQ) showed a delayed but effective response compared to SLD primaquine.

Conclusions: We found marked differences among ACTs and single low-dose 8-aminoquinoline drugs in their ability and speed to block transmission. The findings from this analysis can support treatment policy decisions for malaria elimination and be integrated into mathematical models to improve the accuracy of predictions regarding community transmission and the spread of drug resistance under varying treatment guidelines.

Introduction: Standard meta-analysis of diagnostic test accuracy assumes a binary classification of participants, i.e. diseased or healthy, and estimates the (overall) sensitivity and specificity. Sometimes, however, we need estimates of accuracy for multiple disease states. For example, it is important to know the ability of a test to detect cancer at different stages (early, advanced etc.), since there may be greater potential benefit of diagnosing at an earlier stage, when more amenable to treatment. If sufficient studies report stage-specific sensitivity (“subgroup data”), we might pool these using standard methods. However, often very few studies report this. In a systematic review of the accuracy of tests to detect hepatocellular carcinoma (HCC) among people with cirrhosis, we found that many more studies reported the proportion of detected HCCs that were at each stage (“baseline proportions”), however.

Methods: We propose a method for obtaining meta-analysed accuracy estimates for multiple disease states, combining subgroup and baseline data. Where stage-specific data are not reported, we assume overall sensitivity is an average of stage-specific sensitivities, weighted by the baseline proportions. Study-level random effects allow for heterogeneity and potential between-study correlations among the disease states. We further extend this approach to continuous tests reporting accuracy at multiple thresholds, building on the model of Jones et al, 2019. This produces pooled estimates of accuracy for multiple disease states, at any diagnostic threshold, based on a combination of subgroup and baseline data.

Results: By applying this method to simulated and real datasets from the HCC systematic review, we show that the model can produce results with increased precision compared to those obtained by meta-analysing stage-specific data alone. For one continuous test, alpha-fetoprotein (AFP), only four and five studies respectively reported sensitivity to detect HCCs at a ‘very early’ or ‘early’ stage – the most clinically relevant quantities – and none of these reported sensitivity across multiple thresholds, rendering results from subgroup data alone meaningless. Our model allowed the inclusion of 33 more studies and produced estimates of sensitivity to detect cancer of each stage, and specificity, across all thresholds.

Conclusions: This approach makes the most of descriptive statistics reported in many test accuracy studies to supplement the more informative, but less often reported, subgroup data. This allows us to produce pooled estimates of accuracy for each disease stage or level of severity, across all thresholds.