Presentation (CTI)

OBJECTIVES: Network meta-analysis (NMA) of IPD survival data enables simultaneous comparison of multiple treatments using direct and indirect evidence, with individual-level covariate adjustment and extrapolation beyond trial follow-up. Phillippo et al. (JRSS-A, 2025) proposed a Bayesian survival NMA framework using M-splines to model the baseline hazard, which can accommodate non-proportional hazards by adjusting each spline coefficient per treatment, but this approach is often unstable in our experience. We propose an alternative Bayesian framework using flexible parametric (Royston-Parmar and log-hazard scale) survival models that directly addresses this limitation.
METHODS: We embed flexible parametric survival models within a Bayesian one-step IPD NMA framework, incorporating the treatment network through appropriate parameterisation of baseline hazards and treatment contrasts. Between-trial heterogeneity is captured through random treatment effects. Non-proportional hazards are easily accommodated through simple or complex functional forms of time. Full posterior uncertainty is propagated to all derived quantities via a NUTS sampler. Standardised survival predictions allow population-level inference on survival functions, restricted mean survival times, and treatment rankings. Claude Code (Anthropic) assisted in software development; all AI contributions were reviewed and approved by the authors.
RESULTS: We illustrate using data from five trials of maintenance therapy post autologous stem cell transplant in newly diagnosed multiple myeloma, comparing lenalidomide and thalidomide against placebo on progression-free survival. Individual patient data are available from three trials (McCarthy 2012, Attal 2012, Palumbo 2014), with reconstructed IPD from two further trials (Morgan 2012, Jackson 2019). Flexible parametric models provide smooth survival function estimation across all network nodes and improved fit over standard parametric alternatives, with posterior survival predictions and treatment rankings fully characterised.
CONCLUSIONS: A Bayesian flexible parametric approach to IPD NMA enables robust simultaneous treatment comparison, coherent uncertainty propagation, and clinically interpretable survival predictions. Open-source software implementing all methods is freely available at https://github.com/RedDoorAnalytics.