Validation of a Novel Data-Driven Algorithm to Detect Atypical Prescriptions in Radiation Therapy.
Connor Thropp1,2, Jaroslaw Hepel1,2, Timothy Leech1,2
1Department of Medical Physics, Brown University, Providence, Rhode Island.
Advances in Radiation Oncology
|June 23, 2025
Summary
A new data-driven model effectively identifies atypical radiation therapy (RT) prescriptions across different institutions and cancer types. This validated model enhances patient safety by aiding in the detection of potential RT prescription errors.
Area of Science:
- Medical Physics
- Radiation Oncology
- Health Informatics
Background:
- Erroneous radiation therapy (RT) prescriptions pose significant risks to patient safety, potentially leading to severe injury or death.
- A novel data-driven model utilizing similarity learning was previously developed to identify atypical RT prescriptions within a single institution.
- The prior study's prototype analysis was limited to a single treatment site, necessitating validation across diverse settings.
Purpose of the Study:
- To validate the robustness and generalizability of a data-driven model for identifying atypical radiation therapy prescriptions.
- To assess the model's performance across multiple disease sites (brain and thoracic cancers) and a different healthcare institution.
- To confirm the model's utility in detecting potential prescription errors in varied clinical scenarios.
Main Methods:
- A query of historical RT treatment records from Brown University Health (1995-2021) for brain and thoracic cancer patients was performed.
- Databases were created including RT prescriptions and patient-specific features; simulated anomalies mimicked potential errors.
- The data-driven model was trained and tested using these databases, with performance evaluated using the F1 score.
Main Results:
- High F1 scores were achieved for brain sites: 99% (intensity-modulated RT), 90% (stereotactic RT/radiosurgery), and 94% (3D-RT).
- Comparable F1 scores were observed for thoracic sites: 95% (intensity-modulated RT), 90% (stereotactic RT/radiosurgery), and 95% (3D-RT).
- Statistical analysis confirmed no significant differences between the model's predictions and the ground truth, indicating high accuracy.
Conclusions:
- The model demonstrates feasibility for application across various disease sites and healthcare institutions, confirming its robustness.
- This validated model can serve as a valuable tool for physicians and physicists during peer review chart rounds.
- The model aids in the proactive detection of potential radiation therapy prescription errors, thereby enhancing patient safety.
