Virtual VMAT Psqa Decision Support In Routine Clinical Practice: 3-Month Prospective Results
Abstract
Purpose
To clinically deploy and prospectively evaluate a quantile random forest (QRF) - based “virtual” VMAT patient-specific QA (PSQA) decision-support tool designed to reduce measurement workload while maintaining patient safety.
Methods
15,076 VMAT PSQA measurements from Elekta Agility linacs (2012–2024) were assembled. Plans were generated in Pinnacle and measured using an ArcCHECK diode array (Sun Nuclear) with global gamma analysis (3%/3 mm, 10% threshold). Fourteen plan-complexity features were used, including total MUs, MU factor, average leaf pair opening (ALPO), small aperture scores at 10mm/20mm. Data were randomly split into training (80%) and independent testing (20%); QRF hyperparameters were tuned using 5-fold cross-validation. Decision support used a lower-quantile predicted GPR pass threshold of 90%, while measured PSQA pass was defined as GPR ≥ 95%. Plans were flagged for measurement if: predicted lower-quantile GPR <90%, ALPO <10 mm, total MU ≥2500, or MU factor ≥1000. Unflagged plans were not measured. During the first 3 months of deployment, 700 clinical plans were processed by the QRF model; 70 plans were selected for prospective evaluation with approximately equal representation above and below the 95% measured threshold.
Results
Median prediction accuracy remained high (MAE=0.90%), while lower-quantile MAE was 4.16%, consistent with initial testing and validation. Measured GPR exceeded the predicted lower quantile in 98.6% of cases (69/70); the lone exception measured 91.4% versus predicted 92.2%. All four plans with measured GPR <95% were flagged for measurement. Without flagging criteria, sensitivity and specificity were 87.9% and 25.0%, respectively; incorporating flagging criteria improved specificity to 100% while maintaining sensitivity at 87.9%.
Conclusion
A machine learning-based virtual VMAT PSQA decision-support system was successfully integrated into clinical workflow. Prospective results maintained performance consistent with retrospective validation, enabling safe measurement triage such that ArcCHECK PSQA measurements are currently reserved only for model-flagged plans.