S Luk¹*, P Kalendralis², C Zegers², I Bermejo², A Kalet³, (1) University of Vermont Medical Center, Burlington, Vermont, USA (2) Department Of Radiation Oncology (Maastro), GROW School for Oncology And Reproduction, Maastricht University Medical Centre, Maastricht, The Netherlands (3) University of Washington Medical Center, Seattle, Washington, USA

• 
  S Luk

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

SU-F-BRA-5 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Ballroom A

Purpose: Physics plan and chart review is a key component of quality control processes for detecting high severity incidents in radiation oncology. To assist this process, we proposed an error detection Bayesian network (EDBN) trained with single institutional data to assist initial physics plan review. The EDBN was shown to be effective on detecting potential errors in treatment plans. In this study, we further study the effectiveness of the EDBN with multi-institutional data and aim to improve the generalizability of the model in different clinical settings.

Methods: Treatment data is extracted from the oncology information systems (OIS) of the participating institutions, including University of Washington (UW), Maastro, and University of Vermont Medical Center (UVMMC). The study has three steps. First, the published EDBN trained with UW data is applied directly on Maastro plans with simulated errors embedded as an independent external validation. Then, the EDBN model is modified and the clinical data is processed and harmonized to adapt the differences between as the participating institutions to improve effectiveness and generalizability of the model. Lastly, the modified EDBN is retrained with data from all three institutions and the effectiveness is studied with plans with simulated errors.

Results: Independent external validation of the UW-trained EDBN using Maastro data with simulated errors has shown a decrease in area under the receiver operating characteristic curve from 0.89 to 0.68. To improve the performance, the structure of the EDBN is modified and clinical data from participating institutions are processed and standardized for retraining of the modified EDBN.

Conclusion: The EDBN is less effective in detecting potential errors in another institution that has different clinical practice standards and technologies. A more generalized model, data harmonization and federated learning would be required to improve the performance and make it clinically available in multiple clinics.

Keywords

Quality Assurance

Taxonomy

IM/TH- Formal Quality Management Tools: General (most aspects)

Contact Email