Comparison of statistical machine learning models for rectal protocol compliance in prostate external beam radiation...

Area of Science:

• Radiation Oncology
• Medical Physics
• Machine Learning in Healthcare

Background:

• External beam radiation therapy (EBRT) for prostate cancer necessitates limiting rectal dose to prevent injury.
• Hydrogel spacers can improve rectal sparing, but predicting dose compliance remains a challenge.
• Identifying patients at higher risk of rectal injury pre-treatment is crucial for optimizing EBRT planning.

Purpose of the Study:

• To evaluate the efficacy of statistical machine learning models in predicting rectal dosimetry compliance before EBRT simulation.
• To compare the performance of logistic regression, classification and regression trees, and random forest models.
• To assess the impact of different classification thresholds on model performance for rectal dose outcomes.

Main Methods:

• Treated 176 prostate cancer patients with conventionally fractionated EBRT (74-78 Gy).
• Quantified rectal dose-volume histogram data (V50%, V83%, V96%, V102%) and classified patients based on tolerance thresholds.
• Compared logistic regression, classification and regression tree, and random forest models using AUC, sensitivity, specificity, PPV, and NPV across 50%, 10%, and optimal thresholds.

Main Results:

• Logistic regression achieved the highest AUC (0.844) at the V83% dose level; random forest performed best at V96% (AUC=0.733).
• A conservative clinical threshold of 10% maximized sensitivity for V83% and V96% across all models.
• Logistic regression and random forest models demonstrated good discriminative ability, especially at higher dose levels and with a 10% threshold.

Conclusions:

• Statistical machine learning models are effective for predicting rectal protocol compliance in prostate cancer EBRT planning.
• Logistic regression and random forest models show strong performance in predicting rectal dose outcomes.
• A conservative clinical threshold enhances model sensitivity, confirming the value of these ML approaches over classification and regression trees.