Natural language processing and machine learning to assist radiation oncology incident learning

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Radiation Oncology Safety

Background:

• Incident Learning Systems (ILS) are crucial for improving safety in radiation oncology.
• Manual classification of incidents in ILS can be time-consuming and prone to errors.
• Semi-automation of incident classification can enhance the efficiency and accuracy of ILS.

Purpose of the Study:

• To develop a Natural Language Processing (NLP) and Machine Learning (ML) pipeline for semi-automating incident classification within an ILS.
• To create ML models that recommend labels for three key data elements in the Canadian NSIR-RT taxonomy.
• To improve the usability and efficiency of radiation oncology incident learning.

Main Methods:

• Processed over 6000 incident reports using NLP techniques.
• Trained and evaluated over 500 multi-output ML algorithms using expert-generated labels.
• Identified and tuned the top three models for 'process step', 'problem type', and 'contributing factors'.

Main Results:

• MultiOutputRegressor extended Linear SVR models demonstrated the best performance.
• Models achieved high accuracy in ranking the most appropriate labels for incident classification.
• Achieved label rankings of 1.48 ± 0.03 for process step, 1.73 ± 0.05 for problem type, and 2.66 ± 0.08 for contributing factors.

Conclusions:

• Developed effective NLP-ML models for incident classification in radiation oncology.
• Models will be integrated into an ILS to provide semi-automated label recommendations via a drop-down menu.
• The developed system has the potential to significantly improve the usability, accuracy, and efficiency of radiation oncology ILS.