Machine learning for normal tissue complication probability prediction: Predictive power with versatility and easy...

Area of Science:

• Radiotherapy
• Medical Physics
• Computational Biology

Background:

• The Lyman-Burman Kutcher (LKB) model is widely used for predicting normal tissue complication (NTCP) after radiotherapy (RT).
• However, the LKB model can exhibit numerical instability and relies solely on generalized mean dose (GMD).
• Machine learning (ML) algorithms present a potential advancement with improved predictive power and fewer limitations.

Purpose of the Study:

• To compare the numerical characteristics and predictive performance of the LKB model against ML models.
• To evaluate the efficacy of ML in predicting radiotherapy-induced toxicity.

Main Methods:

• Both LKB and ML models were developed to predict G2 Xerostomia in head and neck cancer patients post-RT.
• Input features included the dose-volume histogram of parotid glands.
• Model speed, convergence, and predictive power were assessed on an independent training set.

Main Results:

• Global optimization was required for a convergent and predictive LKB model.
• ML models demonstrated unconditional convergence and predictive capability, robust to gradient descent.
• ML models surpassed the LKB model in Brier score and accuracy, while ROC-AUC performance was comparable.

Conclusions:

• ML models effectively quantify NTCP, performing as well as or better than LKB models, even for LKB-suited toxicities.
• ML models offer significant advantages in convergence, speed, and flexibility.
• ML models represent a promising alternative for clinical RT planning decisions.