Knowledge-Based Error Detection in External Beam Physician Orders Using Association Rules

Abstract

The physician orders for external beam radiation therapy (EBRT) are associated with top-level treatment decision parameters including prescription dose, number of treatment fractions, treatment modality, treatment positioning, image guidance, etc. Physician order errors manifest as wrong values of individual parameters or logical inconsistencies between multiple parameters, which are difficult to detect even to human experts without going through many data and documentations. The purpose of this work is to investigate an association rule based approach to error detection in physician orders. The goal is to catch those errors earlier so to avoid the costly re-simulation and re-planning. Clinical physician orders for patients who received EBRT treatments from 2008 to 2017 at author’s institution. A total of 3059 individual single-prescription orders for nine disease sites – brain, breast, lung, pelvic, pelvis, prostate, spine, TBI, extremely – were acquired. Each order includes disease attributes and prescription parameters. Seven disease attributes were considered as conditions - site, tumor stage, nodal stage, metastatic stage, intent, laterality and previous treatment. Errors were detected on the four prescription parameters - total dose, fractions, technique and modality. The a priori algorithm was employed to extract frequent item sets from the historical physician orders. The association rules were generated by arranging items in each frequent item set as antecedent items and consequent items. The active association rules were selected according to their support and confident scores. The error detection tool raises an error flag if a new physician order breaks any active association rules. 10 percent of physician orders were randomly chosen and errors (wrong values in prescription parameters) were added manually for testing the performance of the method. A total of 257 active association rules were selected on average for each individual disease sites. The mean values of true positive and false positive rates of error detection were 92.38% and 10.23% respectively for single-prescription cases of nine disease sites. The wrong value of individual physician order parameters and logical inconsistence between physician order parameters could be detected by applying association rules with high positive rate, which could be further improved by optimizing the association rule discovery algorithm. The association rules are human expert understandable and verifiable, and linked directly to historical physician orders. Association rule discovery algorithm can naturally handle physician orders with missing values existing in more than 50% physician orders. The approach supports incorporation into independent error detection tools for assisting manual double-checks on physician orders. The success of the method here also gives promise to further scaling to include patient setup parameters and all treatment sites.