ESTRO 2024 - Abstract Book

S4576

Physics - Machine learning models and clinical applications

ESTRO 2024

from 1 Gy to 70 Gy at 1Gy step size. To reduce the space dimensions and mitigate the multi-collinearity among the 70 variables we adopted the principal component analysis (PCA) and the first five primary principal components were input into the logistic regression classifier for RP risk prediction. The predictive performance was evaluated with the area under the curves (AUC) of the receiver operating characteristic (ROC).

Results:

Dosimetric variables V20 and MLD followed the normal distribution and V5 the quasi-normal distribution from the Shapiro-Wilk test, showing that they are suitable for the subsequent statistical tests. The one-way ANOVA test showed that V20 and MLD were statistically different but not V5 among any two pairs of three methods (Fig. 1(a)). Similarly, the Mann-Whitney U test yielded clear difference between non-RP2 and RP2 for dosimetric variables V20 and MLD but not V5 again (Fig. 1(b)). Univariate logistic regression revealed that the correlation with RP incidence was in the