ESTRO 2024 - Abstract Book

S4461

Physics - Machine learning models and clinical applications

ESTRO 2024

1 National Center for Oncological Hadrontherapy (CNAO), Clinical Department, Pavia, Italy. 2 Istituto Nazionale di Fisica Nucleare, Sezione di Torino, Torino, Italy

Purpose/Objective:

To build a data-driven normal tissue complication probability (NTCP) model for predicting high grade radiation induced temporal lobe necrosis (TLN) in skull-base tumors treated with proton therapy (PT) at a single institution.

Material/Methods:

We retrospectively collected data of 110 skull-base chordoma (81) and chondrosarcoma (29) patients, treated with curative intent PT (prescription dose: 74 Gy(RBE) and 70 Gy(RBE)) between September 2011 and July 2020 [1]. Doses were delivered with pencil beam scanning (PBS) and a sequential boost scheme [2]. RBE-weighted dose was calculated setting a constant value of 1.1 and TL dose was constrained to D2cc<71 Gy(RBE) in the optimization process [3][4]. To guarantee consistency of the results [5], all plans were recalculated with RayStation v.8B. Each TL dose volume histogram (DVH) was sampled to derive 22 dosimetric variables to be evaluated for the development of a structure wise TLN NTCP model, along with 6 selected clinical variables (gender, age, hypertension, diabetes, number of surgeries, surgical techniques). TLN was assessed with MRI and scored according to the Common Terminology Criteria for Adverse Events version 5.0 scale (CTCAE) [6]. Dosimetric and clinical variables proving a potentially significant relation (p≤ 0.1 at t-test) with G2-TLN (CTCAE score≥G2) were considered for developing the G2-TLN NTCP model. The optimal set of predictors was investigated with the bootstrap-enhanced elastic-net (BE-E-Net) regularized logistic regression [7] (1000 bootstrap samples). For each bootstrap sample, the model’s hyper-parameters were optimized via a 5-fold cross-validation, hence the selected predictors along with the major model characteristics were recorded. The number of variables most frequently selected across the bootstrap samples, while minimizing the combined Akaike information criteria (AIC) and Bayesian information criteria (BIC) values, was chosen for the final NTCP model (N). After multicollinearity check, non-penalized logistic regression models were fitted for each combination of N predictors based on the bootstrapped selection frequency (≥50%). Hence, the final G2-TLN NTCP model with N independent predictors was defined based on the AIC criteria. Finally, model performance was evaluated with the 10 fold cross validated area under the receiver operating characteristic curve (AUROC). 95% bootstrap confidence intervals (CI) were computed and the statistical significance of the AUROC evaluated with the permutation test (1000 permutations). Moreover, model calibration and goodness of fit were assessed with calibration plot and Hosmer Lemeshow test. With a median follow-up of 36 months (9-98 months), 40 (36%) patients developed TLN at any grade, 14 (12.5%) developed G2-TLN, and no TLN≥G3 was registered. In G2-TLN, the median volume was 0.46 cc [0.35 cc; 1.34 cc] and the median average and maximum dose received were 66 Gy(RBE) [64 Gy(RBE); 68 Gy(RBE)] and 80 Gy(RBE) [77 Gy(RBE); 83 Gy(RBE)]. In the dataset, no clinical features showed potentially relevant association with G2-TLN, therefore, only a subset of dosimetric variables (all except for D40cc and D50cc) was input to the BE-E-Net procedure. Across the bootstrap samples, elastic-net regularized logistic regression models with two variables were the most frequently selected, contextually minimizing both AIC and BIC values. Based on selection frequencies the following DVH variables were evaluated: D10cc, D20cc, V70, and V75. V75 and D20cc confirmed a statistically significant association with G2-TLN in the logistic multivariable analysis, and reported the lower AIC value compared with the other 2-variable NTCP models. V75-D20cc non-penalized NTCP model (intercept=-5.50[95%CI: -7.57—3.45], and Results: