ESTRO 2024 - Abstract Book

S2626

Clinical - Urology

ESTRO 2024

Material/Methods:

Data have been acquired from Italian centers joining the AIRO-URO initiative. Pts who underwent RP and salvage RT after experiencing BCR have been considered. Data were processed to compute time to the event of distant metastasis development and its status. Only variables with missing values below 15% were considered and chain equation was used for multiple imputation. All participating institution data, except one used as an external test, were included in the train-test. Feature selection employed the Least absolute shrinkage and selection operator (LASSO) method, with non-zero coefficient variables incorporated into the Cox Proportional-Hazards (CoxPH) model. A 5-fold cross-validation was utilized throughout. Performance was assessed using C-index and AUC on the test set. A HR nomogram was constructed from the CoxPH variables.

Results:

A total of 1.651 patients were included in the analysis, divided into a training set of 1.490 (90.2%) patients and an external test set of 161 (9.8%) patients. Initially, 12 variables were considered for LASSO feature selection, but only 4 (pT equal to 3 or 4 at diagnosis, ISUP equal to 4 or 5 at diagnosis, pN1 at diagnosis, and time from surgery to salvage radiotherapy less than 12 months) were incorporated into the CoxPH model. The training set yielded a Chi-squared value of 79.25 with a p < 0.05. The model, when tested on the external set, revealed a C-index of 70.81 and an AUC of 72.20. A nomogram of the model was plotted using the cumulative hazard ratio for distant metastasis as the functional metric (Fig. 1).

Fig. 1 - Nomogram for distant metastasis hazard

Conclusion:

The creation of a nomogram powered by machine learning can optimize the profiling of patients with BCR, guaranteeing precise and timely imaging. After additional refinement and validation, this tool can seamlessly merge into existing clinical procedures.