Impact of daily machine performance check on prediction models for patient‐specific quality assurance of multiple‐target stereotactic radiosurgery
Nicola Lambri, Giacomo Reggiori, Leonardo Nicotera, Veronica Vernier, Carmela Galdieri, Ciro Franzese, Daniele Loiacono, Stefano Tomatis, Marta Scorsetti, Pietro MancosuAbstract
Background
Single‐isocenter multiple‐target (SIMT) stereotactic radiosurgery (SRS) delivered with standard linear accelerators (linacs) achieves high dose conformity through modulated fields with small multi‐leaf collimator (MLC) apertures, making patient‐specific quality assurance (PSQA) a critical component of radiotherapy programs. Complexity metrics are established predictors of PSQA results, while daily quality control checks of linacs represent an additional potential source of predictive information.
Purpose
To investigate whether integrating machine performance check (MPC) parameters with complexity metrics improves machine learning model prediction of the gamma passing rate (GPR) in SIMT SRS PSQA.
Methods
A total of 1054 volumetric modulated arcs from 245 consecutive patients treated with SIMT SRS were included. The median total planning target volume was 5.03 cc (range 0.31–119.65 cc) treated with 10MV‐FFF beams in 1 fraction of 20 or 24 Gy, or in 3–5 fractions of 27–32 Gy. PSQA was performed field‐by‐field using an electronic portal imaging device (EPID) with a gamma criterion of 3%/1 mm and a 95% tolerance limit. Eighteen arc‐specific plan complexity metrics (aPCM) and ten MPC features were considered. Feature selection based on hierarchical clustering of the Spearman's rank correlation matrix yielded fourteen features (five aPCM, nine MPC). Data were split chronologically according to the QA session date, with the earliest 80% of arcs used for training and the most recent 20% for testing. Four histogram gradient boosting models were trained to predict the GPR of each arc: model A (aPCM and MPC), model B (aPCM only), model C (MPC only), and model D (MPC predicting the residuals of model B). Model performance was evaluated using mean absolute error (MAE) and area under the receiver operating characteristic curve (AUC), with bootstrap 95% confidence intervals. Differences between models A and B were assessed with Cohen's f 2 and a Wilcoxon signed‐rank test on residuals. Model explainability was obtained with SHAP (SHapley Additive exPlanations).
Results
Model A showed the best performance (MAE = 1.7% [1.5, 1.9]; AUC = 0.87 [0.79, 0.93]), closely followed by model B (MAE = 1.8% [1.6, 2.0]; AUC = 0.83 [0.75, 0.89]) and model D (MAE = 1.8% [1.6, 2.0]; AUC = 0.83 [0.76, 0.90]). Model C performed markedly worse, particularly in classification ability (MAE = 2.0% [1.7, 2.2]; AUC = 0.64 [0.58, 0.71]). The incremental effect of adding MPC features was small‐to‐medium but not significant on residuals ( f 2 = 0.13 [0.02, 0.27]; Wilcoxon p = 0.213). SHAP identified aPCM as the primary predictors in models A and B, with median MLC gap showing the highest contribution (average absolute SHAP value = 1.66%). Among MPC features, couch rotation accuracy showed the highest contribution ( = 0.27%), considerably smaller than that of aPCM (group mean = 0.07% vs. 0.48%).
Conclusions
MPC features did not provide meaningful complementary information for predicting PSQA results of SIMT SRS. Complexity metrics related to MLC aperture characteristics were the dominant predictors, indicating an association between arc complexity and PSQA outcomes in SIMT SRS. To reduce variability in EPID field‐by‐field PSQA results, we recommend controlling the complexity of SIMT SRS arcs.