CT Radiation Overdose Detection Using Regularized Regression
PI: Yannis Ch. Paschalidis
We are developing a rigorous and automated outlier detection procedure which leverages electronic records for each CT exam to identify all scans with an abnormally high exposure given the characteristics of the patient and the type of the exam. This is a critical problem hospital radiation safety committees are facing, as there is indeed a non-trivial percentage of CT exams that lead to higher radiation dosage than medically necessary, either due to scanner malfunction, human error, or complex interactions between technician settings and scanners’ auto-exposure mechanisms. The approach is based on a novel robust regression methodology we have developed that learns from past CT exam records a relationship between radiation dose and important variables relating to the type of the exam, the volume being scanned, and the type of the scanner. Using such a predictive model, patients whose predicted radiation dose is substantially larger from the radiation dose they actually received are identified as outliers. In an ongoing pilot study at Brigham and Women’s Hospital in Boston, our algorithm performed extraordinarily well in identifying outlying cases, many of which might have been overlooked by a physician based on the current methods of outlier detection.