Feature Engineering for Interpretable Machine Learning for Quality Assurance in Radiation Oncology

Area of Science:

• Medical Physics
• Machine Learning in Healthcare
• Radiology Quality Assurance

Background:

• Chart checking is a time-consuming task for physicists, leading to high cognitive workload.
• Existing automation efforts partially address chart checking but lack data-driven approaches for quality assurance.
• Reducing cognitive load in pretreatment physics chart checks is crucial for efficient and accurate patient care.

Purpose of the Study:

• To evaluate feature selection methods for enhancing machine learning model interpretability and transparency.
• To predict the difficulty of pretreatment physics chart checks using data-driven techniques.
• To reduce the cognitive workload associated with quality assurance processes in medical physics.

Main Methods:

• Comparison of four feature selection methods: chi-square, mutual information, feature importance thresholding, and greedy feature selection.
• Utilized four different machine learning classifiers.
• Employed SMOTE oversampling in conjunction with feature selection techniques.

Main Results:

• Random forest classifier demonstrated the highest performance when combined with SMOTE oversampling and mutual information for feature selection.
• Achieved an accuracy of 84.0%, Area Under the Curve (AUC) of 87.0%, precision of 80.0%, and recall of 80.0%.
• Feature selection methods significantly improved the interpretability and transparency of the predictive models.

Conclusions:

• Data-driven feature selection methods can effectively improve the interpretability and transparency of machine learning models in medical physics.
• The proposed approach offers a pathway to reduce cognitive workload in chart checking processes.
• This study highlights the potential of machine learning to optimize quality assurance in radiation oncology physics.