3.3 Classification Performance of the LR models
ROC curves of LR models based on CMP, NP and CMP+NP were shown inFig. 3 and the corresponding results were illustrated inTable 2 for training set and validation set. When training with
LR classifier, the AUCs of the CMP, NP and CMP+NP models were 0.676
(95%CI:0.546, 0.789), 0.780 (95%CI:0.658, 0.875), and 0.779
(95%CI:0.657, 0.874) in the training set, respectively, which were
confirmed in the validation set by AUCs of 0.788 (95%CI:0.526, 0.944),
0.803 (95%CI:0.543, 0.952), and 0.803 (95%CI:0.543, 0.952),
respectively. Because NP-based skewness filtered by
exponential and squareroot filters
showed meaningful in differentiating CCSK from Wilms’ tumor, the
performance of these features and LR model based on NP was compared by
Delong test, and the results showed there was no significant difference
among them (p>0.05). The Mann–Whitney U-test showed the
distribution of skewness differs between CCSK and Wilms’ tumor. Lower
skewness was observed in Wilms’ tumor, and higher skewness in CCSK.