ESTRO 2021 Abstract Book

S179

ESTRO 2021

Results A total of 3708 X-ray deviations were analyzed and compared between the four different mask systems. Table 1 shows the relatively small deviations between X-ray corrections, which were all in the sub-millimeter range for translations and below 0.8° for all rotations. Kruskal-Wallis-test for translational and rotational deviations grouped by all four mask systems showed a significant difference. Post-Hoc Dunn-Bonferroni-Tests were applied for further classification and lead to significant differences (p<0.001) in lateral and vertical direction between OM and BLOF and between OM and DM. Between the stereotactic mask systems (DM and BLSRS) there was a significant difference in the longitudinal, vertical and yaw direction.

Conclusion The use of the ExacTrac Dynamic enables intrafractional motion detection using the X-ray-based gold standard. As expected, the lowest movement was shown using cranial stereotactic mask systems (DM, BLSRS), which are used in particular for high-precision radiosurgical treatments. However, the evaluation of over 490 treatment sessions has shown a deviation of lower than 0.8mm along all translational directions and lower than 0.8° along all rotational axes using all kind of thermoplastic mask systems with adequate IGRT. Poster highlights: Poster Highlights 10: Lung PH-0270 Prediction models for cardiotoxicity in patients with non-small-cell lung cancer: systematic review F. Tohidinezhad 1 , F. Pennetta 2 , J. Van Loon 1 , D. De Ruysscher 3 , A. Traverso 1 1 Maastricht University Medical Centre, Department of Radiation Oncology (Maastro Clinic), Maastricht, The Netherlands; 2 Maastricht University Medical Centre, Department of Radiation Oncology (Maastro Clinic), Maastrciht, The Netherlands; 3 Maastricht University Medical Center, Department of Radiation Oncology (Maastro Clinic), Maastricht, The Netherlands Purpose or Objective Although lung cancer is still the first leading cause of cancer death, the prognosis has improved leading to more patients at risk to develop cardiac complications. Data-driven evidence of cardiac risk factors would help patient care. This study was aimed A) to review the available prediction models for cardiotoxicity after different treatment modalities in Non-Small-Cell Lung Cancer (NSCLC) patients, and B) to evaluate the most common machine learning techniques and related risk of biases. Materials and Methods A systematic search of MEDLINE (via PubMed) was performed on 10-02-2021 using a Boolean combination of appropriate truncation and indexing terms related to "NSCLC", "prediction models", "cardiotoxicity", and "treatment modalities" (i.e. chemoradiation, particle therapy, immunotherapy, targeted therapy, or surgery). Following exclusion criteria were applied: sample-size of less than 100, no significant predictors in multivariate analysis, lack of model specifications, and case-mix studies. The quality assessment was performed using 13 methodological considerations, including hyper-parameter tuning, test of interaction terms, feature selection technique, and applying shrinkage/penalization methods (Andaur Navarro CL et al. PMID: 33177137). Results Of the 3,056 papers retrieved, 28 prediction models (7 after chemo-radiotherapy, 1 after immunotherapy, and 20 after surgery) were identified. The median sample size and event rate were 288 (range: 109–34,209) and 12.8% (range: 3.9%–49.6%), respectively. As shown in Table 1, 41 distinct predictors were entered in the prediction models, including age (n=15, 54%), history of cardiovascular diseases (n=13, 46%), atrial volume measures (n=11, 39%), gender (n=7, 25%), heart dose (n=5, 18%), and heart volume (n=5, 18%). The risk of developing atrial fibrillation, arrhythmia, and pericardial effusion were predicted in 12, 4, and 2 studies, respectively. Two studies included elderly patients and logistic regression (n=19, 68%) was the most frequent modeling technique. Only five studies reported the discrimination power of the prediction model using the area under the receiver operating characteristic curve (range: 0.57–0.83). The median quality score was 4 and only one study considered 7 model building instructions (Figure 1).