Causal machine learning reveals age-dependent radiation dose effects on mandibular osteoradionecrosis

Area of Science:

• Oncology
• Radiotherapy
• Machine Learning
• Causal Inference

Background:

• Distinguishing causality from correlation is crucial for translating clinical research findings into effective treatments.
• Osteoradionecrosis (ORN) is a significant complication of radiotherapy for head and neck cancers.
• Current treatment guidelines often lack personalized approaches based on individual patient responses.

Purpose of the Study:

• To establish causal relationships between radiation dose parameters and mandibular osteoradionecrosis (ORN) using causal machine learning.
• To identify patient subgroups with varying sensitivities to radiation dose regarding ORN development.
• To explore the potential for personalized radiotherapy treatment planning to reduce ORN risk.

Main Methods:

• Application of causal machine learning, specifically generalized random forests, to retrospective data from 931 head and neck cancer patients.
• Analysis of volumetric-modulated arc therapy (VMAT) radiation dose parameters.
• Integration with explainable machine learning to assess treatment effect heterogeneity across patient demographics, particularly age.

Main Results:

• All investigated dosimetric factors demonstrated significant positive causal effects on ORN development.
• Average treatment effects for dosimetric factors ranged from 0.092 to 0.141.
• Substantial heterogeneity in treatment effects was observed, with patients aged 50-60 showing the strongest dose-response relationships (up to 0.229), while those over 70 showed minimal effects.

Conclusions:

• Radiation dose parameters causally influence ORN risk, with varying effects across age groups.
• Age-stratified treatment optimization and personalized planning for dosimetric factors can potentially reduce ORN.
• Causal inference methods offer a powerful framework for deriving personalized treatment recommendations from retrospective clinical data in oncology and beyond.