Exploring the Feasibility of Network Meta-Analysis for Predictive Models

Despite the fact that Network Meta-Analysis (NMA) has been shown to be a powerful tool in the evaluation of diagnostic tests and treatment comparisons, its applicability to predictive modelling remains underexplored. There is a paucity of studies incorporating NMA into predictive models, mainly due to a lack of published protocol and frameworks. This thesis attempts to address this issue by performing a scoping review to determine the potential of NMA within the context of predictive modelling and the introduction of a novel framework called NetPredict. A literature review revealed seven studies that employed NMA in predictive models. These studies uncovered the absence of clear defined methods and the lack of NMA assumptions. Moreover, two methodologic papers sought to establish the groundwork for the use of NMA in predictive modeling, but were lacking in practical suggestions for transitivity, heterogeneity and model evaluation. The more important assumptions for NMA, like transitivity and consistency were often ignored in all applied papers, which is problematic with regards to the credibility of indirect comparisons. While heterogeneity was examined in studies by the statistic indices of I2 and τ2, few analyses went deeper into effect modifiers than simple descriptive averages. To address these issues, this thesis presented NetPredict, the first methodological framework proposed for NMA synthesis which combines direct and indirect prediction models. Unlike traditional NMAs that depend on treatment effect estimates, NetPredict uses area under the curve (AUC) of the model while maintaining interpretability for heterogeneity and inconsistency measures. Unlike previous approaches that rely on a common comparator, NetPredict employs a flexible grouping mechanism to uphold transitivity and ensure meaningful comparisons across diverse models. By addressing key methodological challenges and refining NMA for predictive modeling, NetPredict paves the way for standardized and reliable comparisons of prediction models across various clinical fields.In that way contributing to the advancement of evidence synthesis in predictive analytics.