Room: Davidson Ballroom A
Purpose: To demonstrate the feasibility of using machine learning to predict results of VMAT patient-specific QC measurements based on features that describe treatment plan complexity and linac performance.
Methods: We used a series of features describing VMAT treatment plan complexity (eg modulation scores, field shape irregularity, leaf gap distributions) and linac performance (eg output, flatness, MLC positional accuracy) to train a support vector machine classification model. Classes were defined as plans that resulted in normal delivery during patient-specific QC (median dose deviation between measurement and treatment planning system calculations within 1%), cold delivery (median dose deviation <-1%), and hot delivery (median dose deviation >1%). A total of 1620 plans were included, with each being measured on one of seven matched linacs using one of three ScandiDos Delta4 diode arrays over a period of approximately one year. Training was performed on 75% of the samples, while 25% of the samples were reserved as an independent test set to assess model performance using a multiclass receiver operator characteristic (ROC) curve analysis.
Results: Testing of the model on an independent data set resulted in areas under the ROC curves of 0.91 for the cold plan class, 0.81 for the normal plan class, and 0.91 for the hot plan class. The features of highest importance in model training included those that describe aperture irregularity, degree of modulation, leaf gap sizes, off-axis aperture positions, linac output, and MLC QC results.
Conclusion: We have demonstrated the potential to predict the results of VMAT patient-specific QC measurements using plan complexity metrics and linac QC results. The ability to accurately predict patient-specific QC results has the potential to allow clinical physicists to deal with failures proactively, rather than reactively, and reduce delays in patient treatment.