Training and temporally validating an NTCP model of acute toxicity after whole breast radiotherapy, including the impact of advanced delivery techniques

Purpose: The aim is to train and validate a multivariable Normal Tissue Complication Probability (NTCP) model predicting acute skin reactions in patients with breast cancer receiving adjuvant Radiotherapy (RT). Methods and materials: We retrospectively reviewed 1570 single-institute patients with breast cancer treated with whole breast irradiation (40 Gy/15fr). The patients were divided into training (n = 878, treated with 3d-CRT, from 2009 to 2017) and validation cohorts (n = 692, treated from 2017 to 2021, including advanced RT techniques). In the validation cohort, patients were classified according to the delivery techniques into static (n = 404) and arc techniques (n = 288). Several clinical/technical information and DVHs of the "skin" (5 mm inner expansion from the body contour) were available. Skin toxicity was assessed during follow-up using the RTOG scale criteria. A multivariable logistic regression model was generated combining skin DVH and clinical parameters, using cross-validation methods that ensured high internal consistency and robustness. The performance of the model was tested in the validation cohort. Results: 14.0 %/17.4 % of patients developed >= G2 toxicity, in the training/validation cohorts, respectively. The resulting multivariable logistic model included axillary lymph node dissection (OR = 1.58, 95 %CI = 1.01-2.48, p = 0.045), hypertension (OR = 1.54, 95 %CI = 1.04-2.27, p = 0.030) and skin V20Gy (OR = 1.008, 95 %CI = 1.004-1.013, p < 0.0001). The AUC of the model was 0.64/0.59 in training/validation, with better performance in the validation cohort if considering only V20Gy (0.62). The model showed satisfactory agreement between predicted and observed toxicity rates: in the validation group, the slope of the calibration plot was 0.96 (R-2 = 0.6) with excellent goodness-of-fit (Hosmer-Lemeshow p-value = 0.99). Looking at each of the three predictors individually, only the role of V20Gy was confirmed in the validation group. Results were similar when considering patients treated with static or arc techniques. Conclusion: An NTCP model for acute toxicity after moderately hypofractionated breast RT was trained. The model underwent temporal validation even for patients treated with advanced delivery techniques. Despite clinical differences and techniques, the confirmation of the dosimetry parameter in the validation cohort highlights its robustness and corroborates the hypothesis that skin DVH may assess the risk with the potential for improving plan optimisation.