ESTRO 2024 - Abstract Book

S795

Clinical - CNS

ESTRO 2024

months). The final training ML models included 870 samples with 85 LF events and 39 binary variables, which generated 200 trees, each with a maximum node size of 15 (Representative tree demonstrated in Figure 1). The OOB performance error was 0.24 for the LF model. Progressing extracranial disease, lower KPS, stable or progression in size of the lesion at first follow-up imaging, and prescription dose were the most important predictors of duration of local control (Figure 2). The OOB performance error was 0.44 for the RN model. Use of targeted therapy post radiation, higher number of lesions treated, lower KPS, and lesion size of 1-1.5 cm were the most important factors predicting the development of RN.

Figure 1: A representative tree in the analysis of LF

Figure 2: The plot shows the importance of each variable for LF with death as a competing risk (event 1 = LF and event 2 = death).