WE‐D‐204B‐05: Online Monitoring and Error Detection of Real‐Time Tumor Displacement Prediction Accuracy Using Statistical Process Control

Purpose : To investigate two statistical process control (SPC) metrics, the Hotelling (T 2 ) statistic and the input‐variable‐squared‐prediction‐error (Q (x) ), for predicting degradation in real‐time tumor displacement accuracy without explicit measurement of tumor displacement. Method and Materials : Independently but concurrently localized tumor and external surrogate positions from a database of Cyberknife Synchrony™ cases (130 treatment fractions from 63 lung tumors, 10 fractions from 5 liver tumors, and 48 fractions from 23 pancreas tumors) were analyzed. Each fraction consisted of 40–112 measurements obtained at an average rate of 0.018 Hz. The first 10 measured internal/external samples in each fraction were used to create fraction‐specific models of tumor displacement using external surrogates. The regression coefficients relating the 3D positions of the 3 skin markers to the 3D tumor positions were calculated using partial‐least‐squares (PLS) regression. The PLS model was applied to all subsequent localizations in the fraction. The T 2 ‐ and Q (x) ‐statistics in the training data were used to develop 90 th , 95 th and 99 th percentile ranges of expected T 2 and Q (x) values. The sensitivities and specificities of T 2 , Q (x) , T 2 ̆Q (x) , and T 2 ̑Q (x) for predicting real‐time tumor displacement errors greater than 3mm and 5mm were determined. Results : The T 2 , Q (x) , T 2 ̆Q (x) , and T 2 ̑Q (x) statistics' sensitivities and specificities varied with error threshold and acceptable percentile ranges of values. In general, the Q (x) statistic was associated with high sensitivity and low specificity, while the T 2 statistic was associated with moderate sensitivity and moderate specificity. For 90 th percentile T 2 , 99 th percentile Q (x) and 3 mm error, the sensitivities of T 2 , Q (x) , T 2 ̆Q (x) , and T 2 ̑Q (x) were 69%, 88%, 92%, and 64%, respectively, and the specificities were 62%, 37%, 28%, and 72%, respectively. Conclusion : This study illustrates the feasibility of SPC metrics for detecting breakdowns in tumor displacement prediction accuracy using external sensors. Conflict of Interest : Funded by NIH grant CA 124766