Diabetes Mellitus is a chronic metabolic disorder requiring accurate early detection. While machine learning is widely applied in diabetes prediction, the degree of feature ranking agreement across algorithm variants remains insufficiently quantified. This study evaluates the pairwise rank consistency of feature hierarchies generated by five Relief-Based Algorithm (RBA) variants ReliefF, SURF, SURF*, MultiSURF, and MultiSURF* on the Pima Indians Diabetes Dataset (N = 768, 8 predictor attributes, class distribution: 500 negative and 268 positive instances). To prevent target-based data leakage during preprocessing, medically implausible zero values in clinical variables were imputed using target-independent global medians prior to Min-Max normalization. Feature weighting was executed via the skrebate framework, followed by pairwise rank consistency analysis using Kendall’s Tau-b correlation coefficients alongside exact p-value tests and concordant/discordant pair counts. Empirical findings reveal a structural divergence in ranking behavior: non-star variants (SURF and MultiSURF) exhibited a strong rank agreement with each other ( τ = 0.8571, p = 0.0017, c = 26, d = 2), while maintaining high alignment with baseline ReliefF ( τ = 0.7857, p = 0.0055). Conversely, star variants (SURF* and MultiSURF*) yielded orthogonal ranking structures relative to non-star methods ( τ = 0.0000, p = 1.0000, c = 14, d = 14). These empirical observations demonstrate that choice of RBA neighborhood formulation substantially alters the resulting attribute rank order on this dataset representation, emphasizing the importance of accounting for algorithm-specific ranking mechanics when selecting feature subsets.