Bayesian network models for error detection in radiotherapy plans.

Area of Science:

• Medical Physics
• Artificial Intelligence
• Oncology

Background:

• Radiotherapy plan errors pose risks to patient safety.
• Current verification methods may not be fully comprehensive.
• Probabilistic models offer a novel approach to error detection.

Purpose of the Study:

• To design and develop a probabilistic network for detecting errors in radiotherapy plans.
• To integrate this network into initial plan verification processes.
• To enhance the accuracy and efficiency of radiotherapy plan quality assurance.

Main Methods:

• Developed Bayesian networks modeling radiotherapy plans.
• Interviewed domain experts to define network topology and interdependencies.
• Utilized Hugin Expert software and clinical data to populate conditional probability tables.
• Trained networks on 4990 de-identified prescription cases over 5 years.

Main Results:

• Achieved high accuracy in error detection for lung (AUC 0.88), brain (AUC 0.98), and female breast cancer (AUC 0.89) networks.
• The brain network demonstrated superior performance compared to human experts (AUC 0.90 ± 0.01).
• Networks effectively flagged potential errors in test scenarios with 1.5% introduced error rates.

Conclusions:

• Probabilistic models, specifically Bayesian networks, are feasible and effective for radiotherapy plan error detection.
• These models can serve as valuable decision support tools in radiotherapy plan verification.
• The developed networks show potential for improving patient safety by reducing radiotherapy errors.