Prediction Model for Defects in Lead and Lead-free Aprons

Abstract

Abstract Personal radiation protective equipment (PRPE) is prone to defects in the attenuating layers, resulting in inadequate protection. Hence, quality control (QC) of PRPE is needed to assess its integrity. Unfortunately, QC of PRPE is laborious and time consuming. This study aimed to predict the QC outcome of PRPE without x-ray imaging based on readily available predictors. PRPE QC data of a general hospital from 2018 to 2023 was used for both prediction models based on logistic regression and random forests (RF). The data were divided into a training set containing all data from 2018 to 2022 and a holdout set containing the data from 2023. The predictors were brand, age, size, type, visual defects, and department. The prediction performances were compared using confusion matrices and visualized with receiver operating characteristic (ROC) curves. Prediction accuracies of at least 80% were achieved. Further model tuning especially improved the RF model to a precision up to 97% with a sensitivity of 80% and specificity of 86%. All predictors, except visual defects, significantly impacted the probability of passing. The predictor brand had the largest contribution to the predictive performance. The difference in pass probability between the best-performing and the worst-performing brand was 35.1%. The results highlight the potential of predicting PRPE QC outcome without x rays. The proposed prediction approach is a significant contribution to an effective QC strategy by reducing time consuming x-ray QC tests and focusing on garments with higher probability of being defective. Further research is recommended.