ESTRO 2024 - Abstract Book

S4459

Physics - Machine learning models and clinical applications

ESTRO 2024

Material/Methods:

We performed a pooled retrospective and prospective analysis including all patients treated for spinal metastases at our Charité CyberKnife Center between 2011 and 2021. CyberKnife treatments were performed in 1-5 fractions using a non-isocentric treatment approach. The local control was accessed using the SPINO criteria. For all dosage parameters, the biological equivalent dose was calculated using alpha/beta = 10 for tumor, alpha/beta = 2 for spinal cord and alpha/beta = 3 for bony tissues. For the predictive models, we evened out our dataset using random over sampling (3), to generate an even distribution concerning the events. A separate simulation was performed for each method. A 75/25 % - split ratio was used to generate the training set on which the tuning was performed, and the testing set. Due to the limiting small number of events and to achieve a comparability between the different approaches, we only used the two parameters “median PTV-dose” and “median vertebral-body-dose” as the independent variables in all our models. To access the performance of our models more robustly, we generated 1000 further testing sets using repeated random oversampling. All statistical analyses were performed using the software SPSS 29 and R, particularly the “kernlab” (4), “ROSE” (5), “mlr3” (6), “performance” (7), “caret” (8) and “lmtest” (9) packages. A total of 121 patients with 235 spine metastases were included in this study. The most common primary tumors were prostate cancer (44.3 %), breast cancer (14.9 %) and renal cell carcinoma (11.5%). Most of the lesions were treated in a single session (151, 64.3%) rather than hypofractionated (84, 35.7%). For single session treatments, the median prescribed dose was 20 Gy (15 – 22 Gy) to a median PTV of 7.64 ccm (0.19 – 173.67 ccm). For lesions treated with hypofractionated SBRT, the median prescribed dose was 24 Gy (19.5 – 30 Gy) to a median PTV of 31.88 ccm (1.48 – 154.52 ccm). During a median follow up of 16.5 months (3.0 – 84.5 months), 44 lesions (23.0%) progressed locally and 37 lesions (19.4%) experienced complications with the most common being vertebral body fractures (25, 67.6%). Logit models were fitted for our binary endpoints “local failure” and “vertebral body fractures”. For local failure, the model did not fit (Hosmer-Lemeshow test: p = .005) and showed no better fit compared to the null model (log likelihood ratio test: p = .198). However, the logit model for vertebral body fractures provided both a good overall fit (Hosmer-Lemeshow test: p = .612) and a better fit than the null model (log-likelihood ratio test: p < .001). The overall prediction of adverse outcomes had an average accuracy of 56.57% (95% CI: 56.21% – 56.92%) for local recurrences and 43.84% (95% CI: 43.48% – 44.19%) for fractures. The positive predictive values were quite low, reaching only 58.62% (95% CI: 58.12% – 59.12%) for the failures and 44.91% (95% CI: 44.61% – 45.22%) for the fractures. SVM models were then fitted for both the outcomes. Primarily, a kernel-based machine using the Laplacian kernel function as a soft edge classifier was chosen. Using a random search tuning algorithm with 300 iterations optimizing the overall accuracy, resulted in the hyperparameters σ1 = 92, cost for misclassification C1 = 38.9 for local failure and σ2 = 88.8, C2 = 33.3 for vertebral body fractures (k2). Those classifiers resulted in an accuracy of 100% for both classifiers in the training sets and an average accuracy of 89.25% (95% CI: 89.01% – 89.50%) and 95.39% (95% CI: 95.22% – 95.57%) respectively in the generated test sets. Also, the average positive predictive values improved over the logit models for local failure (96.61%, 95% CI: 96.33% – 96.88%) and for vertebral body fracture (99.59%, 95% CI: 99.48% – 99.70%). Results:

Conclusion: