Encrypted login | home

Program Information

A Method to Detect Radiation Therapy Physician Order Errors Using Bayesian Networks

X Chang

X Chang1*, H Li1 , A Kalet2 , D Yang1 , (1) Washington University in St. Louis, St. Louis, MO, (2) University of Washington, Seattle, WA


TU-FG-702-2 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 702

Purpose: To investigate the ability of a machine-learning based approach to detect errors in radiation therapy physician orders.

Methods: EBRT physician orders from 2007 to 2015 were obtained from the treatment management system at author’s institution. A total of 5374 individual orders for 14 disease sites were acquired (3772 single-prescription, 400 concurrent boost and 401 sequential boost cases). Each order includes seven disease attributes (site, tumor stage, nodal stage, metastatic stage, intent, laterality and previous treatment) and four prescription parameters (total dose, fractions, technique and modality). A Bayesian network model was developed for each of the three prescription types. Each model was designed in three layers: disease attributes, order parameters and an anomaly flag. A Bayesian learning method with Dirichlet prior was employed to train the models. The conditional joint probabilities of the anomaly flag and the prescription parameters were employed to detect errors in the prescription parameters with respect to the disease attributes. 10 percent of the physician orders were randomly chosen, and prescription parameter errors were created manually for testing the performance of the method.

Results: The mean values of true positive and false positive rates of error detection were 93.17% and 9.69%, respectively for single-prescription (14 disease sites), 91.77% and 8.85% respectively for concurrent boost (3 sites), and 91.63% and 10.35% respectively for sequential boost cases (4 sites).

Conclusion: Prescription parameter errors can be detected by the trained Bayesian network models with high positive rates. The approach supports further development of incorporating a Bayesian network model into error detection tools for assisting manual checks of physician orders. Use of automated validation for physician order-entry is a viable approach to improve patent safety in radiation oncology and could have application to other clinical domains which heavily use electronic medical records systems.

Funding Support, Disclosures, and Conflict of Interest: Funding: AHRQ R01-HS022888; No conflict of interest; Disclosures: Authors have technology licensing fee from Viewray

Contact Email: