Causal Machine Learning Analysis of Empirical Relative Biological Effectiveness (RBE) for Mandible Osteoradionecrosis...

Area of Science:

• Oncology
• Medical Physics
• Machine Learning

Background:

• Osteoradionecrosis (ORN) of the mandible is a severe complication of head and neck (H&N) cancer radiotherapy.
• Traditional statistical models often struggle to establish causal relationships between treatment factors and ORN.
• Causal machine learning offers a novel approach to investigate these relationships in real-world data.

Purpose of the Study:

• To apply causal machine learning to determine empirical relative biological effectiveness (RBE) for mandible ORN.
• To compare outcomes between pencil-beam-scanning proton therapy (PBSPT) and volumetric-modulated arc therapy (VMAT) in H&N cancer patients.
• To identify dose-volume constraints (DVCs) associated with the highest causal effects for ORN.

Main Methods:

• Included 1,266 H&N cancer patients (335 PBSPT, 931 VMAT).
• Utilized 1:1 case-matching based on propensity scores and standardized mean differences to control for confounding factors.
• Employed causal forest (CF) modeling to investigate causal effects between dosimetric factors and ORN incidence, deriving empirical RBE values.

Main Results:

• After matching, residual confounding persisted, necessitating advanced causal inference.
• CF modeling identified mandible DVCs for ORN, with PBSPT showing lower critical volumes than VMAT.
• Empirical RBE values for PBSPT exceeded 1.1 in the moderate dose range (40-60 Gy), with specific values noted at 40, 50, and 60 Gy.

Conclusions:

• This study demonstrates a novel causal machine learning approach for evaluating ORN risk in radiotherapy.
• Derived empirical RBEs suggest proton therapy's RBE may significantly exceed 1.1 in moderate dose ranges.
• Incorporating variable RBE into PBSPT treatment planning is crucial for mitigating mandible ORN risk.