Figure 6.

Heatmaps on mean model generalizability improvements (A) and statistical test results (B) after feature robustness filtering. Model generalizability is defined as the difference between training and testing AUCs, AUC_testing - AUC_training. A score closer to zero shows better generalizability. In general, model generalizability improved after feature robustness filtering, as shown by the negative values on the heatmaps (A) for both filtering thresholds. Greater improvements were observed with the higher filtering threshold (ICC > 0.95). Moreover, more significant differences are shown by the smaller P-value. However, the predictions of LR on the dataset HN-PETCT showed worse generalizability after feature robustness filtering and the opposite trend of generalizability change and statistical test results with increasing filtering thresholds.