Radiotherapy toxicity prediction using knowledge-constrained generalized linear model.

Area of Science:

• Medical Physics
• Biostatistics
• Oncology

Background:

• Radiotherapy (RT) is crucial for cancer treatment, but dose spillage to normal organs causes toxicity.
• Predicting radiotherapy toxicity relies on Normal Tissue Complication Probability (NTCP) models using dosimetric variables.
• Existing models face challenges with high-dimensional data and limited sample sizes, often lacking domain knowledge integration.

Purpose of the Study:

• To develop a novel statistical model that incorporates medical domain knowledge for improved NTCP modeling.
• To address limitations of data-driven variable selection techniques in radiotherapy toxicity prediction.
• To enhance the interpretability and accuracy of models predicting complications from radiation therapy.

Main Methods:

• Proposed a knowledge-constrained generalized linear model (KC-GLM) integrating domain knowledge via constraints.
• Formulated KC-GLM with non-negativity, monotonicity, and adjacent similarity constraints on model coefficients.
• Developed an equivalent transformation for KC-GLM to enable solutions using standard optimization solvers.

Main Results:

• KC-GLM demonstrated superior variable selection interpretability compared to existing techniques.
• The model avoided counter-intuitive and misleading results often seen in purely data-driven approaches.
• Experiments on simulated and real-world datasets (prostate and lung cancer) showed improved prediction accuracy for KC-GLM.

Conclusions:

• KC-GLM offers a robust framework for building more reliable and interpretable NTCP models.
• Integrating medical domain knowledge significantly enhances the predictive power and clinical relevance of radiotherapy toxicity models.
• This approach holds promise for optimizing radiation therapy planning and minimizing patient side effects.