Model Prediction
The five-times repeated 3-fold cross-validation logistic regression
models with the 2 selected radiomics features yielded a mean AUC value
of 0.879 (SD=0.04) on the training and 0.865 (SD=0.08) on the testing
set (prediction cut off = 0.3) for MYCN amplification status prediction
(Figure 4A). After the first binary outcome identification (i.e. MYCN
amplified versus not amplified), a following analysis considering also
the additional “gain” mutational status was ran. Interestingly out of
the 11 gain patients, nine were located under the amplification
prediction cutoff threshold of 0.3. Figure 4B discloses the gain
patients in the general model. This may support the existence of a
common radiomics pattern between non amplified and gain patients,
confirming the similar clinical behavior within the two classes. In
order to test the ability of the two selected radiomics features to
predict survival outcomes, actual OS data were plotted with radiomics
based predictions and relative Kaplan-Meier curves were designed.
Log-rank tests have been done between observed and predicted amplified
and not amplified patients curves (p=0.003 and 0.05 respectively), while
no statistical significant difference has been observed between the
observed and actual curves (p=not significant), (Figure 5).