Fig. 6
From: Transfer learning prediction of type 2 diabetes with unpaired clinical and genetic data
Performance distributions across 100 independent runs for five model architectures−XGBoost, plain DNN, intermediate fusion (IF), the graph-based MGCN-CalRF model (GCN), and our transfer-learning approach (TL)−evaluated on the (A) validation and (B) test sets. Each box-and-whisker plot displays the spread of (left to right) \(\hbox {F}_1\)-score, \(\hbox {F}_2\)-score, AUROC, and balanced accuracy. Horizontal brackets mark pair-wise differences quantified with two-sided paired t-tests (ns = not significant; * \(p<0.05\); ** \(p<0.01\); *** \(p<0.001\); **** \(p<0.0001\)).
