The Meta-Analysis Genome: Unsupervised Phenotyping of 307 Cochrane Reviews

Can unsupervised clustering of multi-dimensional quality metrics from eight independent projects reveal natural archetypes among Cochrane systematic reviews? We assembled a 26-feature matrix for 307 reviews common across Fragility Atlas, PredictionGap, outcome reporting bias, entropy, evidence half-life, MetaRepair, EvidenceTribunal, and conformal meta-analysis projects, standardized with z-score normalization. Principal component analysis provided dimensionality reduction, followed by K-means clustering into four groups matching quality archetypes with ANOVA F-tests identifying the most discriminating features. Four clusters emerged with median prevalence of pathological reviews at 24 percent with 95% CI 19 to 30: Healthy, Moderate, Pathological requiring over 60 percent correction, and Mixed. Robustness score and adversarial gap were the two most discriminating features with F-statistics exceeding 50 across all four identified clusters. This genome approach provides a data-driven taxonomy beyond single-metric quality assessment toward multidimensional evidence phenotyping. The limitation is that the taxonomy depends on specific input projects, and adding or removing a data source would alter cluster boundaries.