Random forests to predict rectal toxicity following prostate cancer radiation therapy

Area of Science:

• Radiation oncology
• Medical physics
• Machine learning in healthcare

Background:

• Predicting late rectal toxicity is crucial for optimizing radiation therapy for prostate cancer patients.
• Traditional normal tissue complication probability (NTCP) models have limitations in accuracy.

Purpose of the Study:

• To develop and validate a random forest NTCP (RF-NTCP) model for predicting rectal toxicity post-prostate cancer radiotherapy.
• To compare the performance of the RF-NTCP model against established NTCP models.

Main Methods:

• Collected clinical data and dose-volume histograms (DVH) from 261 prostate cancer patients.
• Trained a random forest model to predict 5-year rectal toxicity and bleeding.
• Compared RF-NTCP model performance against the Lyman-Kutcher-Burman (LKB) model using area under the receiver operating characteristic curve (AUC).

Main Results:

• The RF-NTCP model demonstrated predictive capabilities for overall rectal toxicity and bleeding.
• RF-NTCP achieved AUC values ranging from 0.66 to 0.76.
• The LKB model showed statistically significantly inferior performance with AUCs ranging from 0.62 to 0.69.

Conclusions:

• The RF-NTCP model is a potentially valuable tool for predicting late rectal toxicity.
• This model incorporates factors beyond DVH, offering a competitive alternative to classic NTCP models.