ESTRO 2024 - Abstract Book

S4543

Physics - Machine learning models and clinical applications

ESTRO 2024

ACUG2+ in the training population was 14.4% (13.2% G2, 1.2% G3). Univariate analysis found a significant correlation with PTV volume, hypertension, diabetes, obesity, axillary lymph node dissection (ALND), and skin dose parameters in the range V5-V34Gy. V20Gy (representing 50% of the prescribed dose) was chosen as the optimal DVH predictor based on the selection procedure. The resulting multifactorial model included V20Gy (OR=1.008, p<0.0001), hypertension (OR=1.54, p=0.030), and ALND (OR=1.58, p=0.045). The AUC of the model was 0.64 (0.60-0.67, p<0.0001). The resulting dose-response curves against skin V20Gy are shown in Figure 1 for patients with none, one or two risk factors.

In the validation cohort, 17.4% of patients exhibited acute skin reactions (16.1% G2 and 1.3% G3). Static and arc treatment reported a similar rate of toxicity (17.2% vs 17.7% with a p = 0.8). The model satisfactorily predicted the validation data (p<0.0001) (see Figure 2): the calibration plot for all patients showed a slope of 0.90 and an intercept of 0.068 (R2=0.71). Static techniques showed a slope of 0.93 and an intercept of 0.047 (R2=0.98), while arc techniques had a slope of 0.92 and an intercept of 0.069 (R2=0.62).