ESTRO 2024 - Abstract Book

S5156

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Material/Methods:

Data of 330 consecutive esophageal cancer patients treated with trimodality therapy (neoadjuvant chemoradiotherapy followed by surgery) at our institution between 2003 and 2021 was included. Postoperative pulmonary and cardiac complications 90 days after surgery were analyzed. Logistic regression modeling was performed at every step in time, i.e., stepwise adding every individual patient treated chronologically. Clinical (age, gender, BMI) and dosimetric (mean lung dose (MLD), mean heart dose (MHD)) features were used along with treatment-related variables capturing 2 gradually implemented major treatment evolutions (open vs minimal invasive esophagectomy (MIE) surgery, 3D conformal radiotherapy (3DCRT) vs intensity modulated radiotherapy (IMRT)). An initial NTCP model was developed in the first 150 patients (treatment period 2003-2011). Three NTCP model adaptation strategies (S) were explored at each step in time, S1: the initial model was kept unchanged (but parameters refitted to the whole retrospective dataset available), S2: a new model was built based on the whole retrospective dataset, and S3: a new model was built only based on the 150 most recently treated patients. For each strategy, the NTCP model performance (AUC and calibration slope) was assessed in the patient data of the future (50 next patients). The inclusion of MLD and MHD features was enforced (with positive regression coefficients) in every pulmonary and cardiac NTCP model, respectively. A systematic decrease in postoperative pulmonary complication rate was observed in time, from 48% during initial model building to 22% during final model validation. The cardiac complication rate was fluctuating between 25% and 45%. Large variability of NTCP model performance throughout time was observed with all 3 strategies. For pulmonary complications (Figure 1) the initial model (S1: MLD risk factor) immediately showed a complete deterioration of its performance (AUC=0.50) when validated in the next 50 patients. Strategies S2 and S3 (implying model adaptation) picked up quickly the variable histology to accommodate this. Near the end of the time range, model performance under S2 deteriorated after it allowed for the inclusion of the surgery technique variable. The risk reduction associated with MIE (consistently included in the S2 and S3 models from patient 258 onwards) could not be validated in the most recent data. The simple MLD model surprisingly was by far the best performing model in this recent data. Average AUC (95%CI) in time was 0.53 (0.48; 0.60), 0.63 (0.59; 0.72) and 0.55 (0.47; 0.66) for S1, S2 and S3, respectively. For cardiac complications (Figure 2) the initial model (S1: age and MHD risk factors) started with a validation AUC around 0.65. S2 confirmed the initial model at most points in time. In data of the last treatment years, the S3 strategy which is most flexible for including recent associations with outcome, for the first time convincingly led to an improved model performance and a significantly earlier inclusion of the surgery technique variable (which effectively proved relevant for the cardiac endpoint at validation in the most recent data). Average AUC (95%CI) in time was 0.63 (0.55; 0.70), 0.63 (0.55; 0.70) and 0.61 (0.51; 0.75) for S1, S2 and S3, respectively. Similar results were observed for slope values. The radiotherapy technique variable was selected in none of the adapted NTCP models. Using less than 150 patients for initial model building or model adaptation under S3 gave worse results. Results:

Conclusion:

Ensuring the accuracy of NTCP model predictions in time proves to be a highly complex forecasting task. None of the tested strategies for continuous NTCP model adaptation proved robust and flexible in all situations of treatment