Prediction and classification of VMAT dosimetric accuracy using plan complexity and log-files analysis

Purpose: We presented different machine learning models based on log files analysis and complexity indexes to predict and classify the dosimetric accuracy of VMAT plans. Methods: A total of 1302 VMAT arcs from 651 treatment plans were analyzed using the modulation complexity score (MCS) and the dynamic log-files generated by the linac. Predicted and measured fluences were compared using gamma-analysis in terms of mean gamma-values (gamma mean) and gamma-pass rate (gamma%). A kernel regression model was developed aiming to predict individual gamma% and gamma mean values. Multinomial logistic regression (LR), Naive-Bayes (NB) and support vector machine (SVM) models were developed based on MCS values to classify QA results as "pass" (gamma%greater than90 % and gamma mean < 0.5), "control" (80 % < gamma% < 90 % and 0.50 < gamma mean < 0.75) and "fail" (gamma% < 80 % and gamma mean > 0.75). Training, validation and testing groups were used to evaluate the model reliability. A complexity-based traffic light protocol was implemented to flag pass (green light), control (orange light) and failed plans (red light). Results: Prediction accuracy of residuals for gamma% was 2.1 % and 2.2 % in the training and testing cohorts, respectively. For 2 %(local)/2mm, both gamma% and gamma mean classification performances reported weighted precision, recall and F1-values greater than 90 % for all machine learning models. The optimal MCS threshold value for the identification of failed plans was 0.130, with a sensibility and specificity of 0.994 and 0.952, respectively. The optimal MCS threshold for reliable plans was 0.270, with a sensitivity and specificity of 0.925 and 0.922, respectively. Conclusions: Machine learning can accurately predict the dosimetric accuracy of VMAT treatments, representing an efficient tool to assist patient-specific QA.