Addressing the Overlooked Inquiry in Feasibility Assessment of Network Meta-Analysis through the Covariate-Optimized Spatial Method for Integrated Clustering (COSMIC)

OBJECTIVES: A feasibility assessment is recommended to evaluate whether the studies included in a Network Meta-Analysis (NMA) share homogeneity in terms of trial and baseline (T&B) characteristics. Simply relying on narrative synthesis might not be enough, especially when dealing with factors that are distributed heterogeneously across trials. Our objective is to introduce a quantitative framework to identify clusters of similar studies and key features responsible for assigning studies to different clusters creating between-cluster variability.

METHODS: A simulated dataset with 25 studies was created, featuring four trial characteristics and eight baseline characteristics. The elbow method, based on the study features, was used to determine the within-cluster sum of squares (WCSS) for different cluster numbers and identify the 'elbow' point, where WCSS stabilizes. Following that, K-means clustering utilizing the Hartigan and Wong (1979) algorithm was applied and adjusted for the available features. This method aims to partition datapoints into k groups, minimizing the sum of squares from datapoints to the assigned cluster center. This algorithm is suitable for smaller datasets due to its iteration over all datapoints. Relative importance analysis (RIA) was conducted to identify key features influencing between-cluster variability and results were presented in percentages.

RESULTS: The Elbow method identified k=3 as the optimal number of clusters. Using K-means clustering with 3 centroids, the cluster plot demonstrated that Cluster 1 had three, Cluster 2 had four, and Cluster 3 had remaining studies. RIA highlighted that Percentage of metastatic patients (24%), Percentage of former smokers (15%), and Percentage of female patients (13%) are the top three features reported in the studies that contribute the most to the determination of the similarity or dissimilarity among trials resulting into three different clusters.

CONCLUSIONS: Clusters of similar studies identified through COSMIC can be used for sensitivity analysis during NMA. It also identifies key features contributing to significant between-cluster heterogeneity.