Fig. 7: Statistical effect size comparisons between model pairs. | Communications Medicine

Fig. 7: Statistical effect size comparisons between model pairs.

From: Volumetric spline-based Kolmogorov-Arnold architectures surpass CNNs, vision transformers, and graph networks for Parkinson’s disease detection

Fig. 7

a Isolated analysis showing effect sizes in AUC comparisons for individual datasets. b Combined analysis of effect sizes across all datasets. c Hold-out analysis effect sizes for generalization assessment. Horizontal bars represent Cohen’s d effect sizes, with blue bars indicating 2D vs. 2D comparisons and red bars indicating 3D vs. 3D comparisons. Significance markers: *p < 0.05, **p < 0.01, * **p < 0.001. Exact p-values and corresponding effect sizes for all comparisons are reported in Supplementary Tables 10–13. Sample sizes: n = 59 independent subjects (PPMI), n = 43 (NEUROCON), n = 40 (Tao Wu) for isolated and hold-out analyses. All comparisons shown are same-dimension comparisons (2D vs. 2D or 3D vs. 3D) with conservative z-test p-values.

Back to article page