Method for Automatic Selection of Parameters in Normal Tissue Complication Probability Modeling.

Area of Science:

• Radiation Oncology
• Medical Physics
• Biostatistics

Background:

• Normal tissue complication probability (NTCP) models are crucial for predicting adverse events after radiation therapy.
• Developing accurate NTCP models often involves complex statistical analysis and expert knowledge.
• Existing methods can be time-consuming and challenging when dealing with numerous potential predictive factors.

Purpose of the Study:

• To introduce a fully automatic method for generating multiparameter NTCP models.
• To compare the performance of the automatically generated model (AGM) with a previously published expert-derived model.
• To assess the predictive accuracy and identify key factors for cystitis in rectal cancer patients.

Main Methods:

• A cohort of 345 rectal cancer patients treated with external radiation therapy was analyzed.
• Principal component analysis (PCA) reduced bladder dose-volume histogram data into 8 components.
• An automated algorithm, employing bootstrap sampling, variance inflation factor, and genetic algorithms, generated an ordinal logistic regression NTCP model (AGM).

Main Results:

• The automatically generated model (AGM) and the published model demonstrated comparable predictive performance on both training and test datasets (P > .05).
• Both models identified similar clinical and dosimetric factors as significant predictors of cystitis.
• While both models showed good explanatory performance on training data, this decreased on the test data (P < .05).

Conclusions:

• The automatically generated NTCP model (AGM) exhibits equivalent predictive value to an expert-derived model.
• This automated approach has the potential to save time and standardize variable selection in NTCP modeling.
• The method is particularly beneficial for handling a large number of parameters in NTCP model development.