Author: Jingyuan Chen, Sheng Li, Tianming Liu, Wei Liu, Zhengliang Liu, Zhong Liu, Daniel Ma, Samir H. Patel, Guangya Wang, Yunze Yang 👨🔬
Affiliation: University of Miami, Mayo Clinic, School of Data Science, University of Virginia, School of Computing, University of Georgia, Department of Radiation Oncology, Mayo Clinic, Institute of Western China Economic Research, Southwestern University of Finance and Economics 🌍
Purpose:
Traditional patient outcome analyses relied heavily on conventional statistical models that primarily elucidate correlation rather than causal relationships. In this study, we aim to identify key dosimetric and clinical variables that causally contribute to the development of osteoradionecrosis (ORN) of the mandible in patients with head and neck (H&N) cancer treated with radiotherapy (RT).
Methods:
This study included 1,266 H&N cancer patients, with 931 patients treated using volumetric-modulated arc therapy (VMAT) and 335 patients treated using pencil-beam-scanning proton therapy (PBSPT). Mandible dose volume histogram metrics (Dmax, Dmean, V40Gy, V50Gy, V60Gy, V70Gy) were extracted. Relative biological effectiveness (RBE) of 1.1 was assumed for patients treated with PBSPT. Clinical variables included age, gender, tumor stage, chemotherapy, hypertension or diabetes, dental extraction history, smoking history, and current smoking status.
We applied the generalized random forests (GRF) method, a non-parametric Causal Machine Learning (ML) approach suited for estimating heterogeneous treatment effects. GRF enabled us to quantify causal influence of dosimetric variables on ORN and to assess the importance of clinical factors. Robustness of the GRF results was evaluated.
Results:
For the VMAT group, our analysis suggested that DVH metrics such as Dmax, Dmean, V40Gy, V50Gy, and V60Gy bear significant causal relationships with ORN and V60Gy emerged as the most influential dosimetric metric, with an estimated treatment effect of 0.185 (95% CI: 0.078–0.292). Among PBSPT patients, V50Gy exhibited the largest estimated treatment effect of 0.161 (95% CI: -0.080–0.440). Age, tumor stage, and hypertension were identified as the most influential clinical variables contributing to ORN. Specifically, patients aged near 57–65 years appeared to be at a heightened risk for both VMAT and PBSPT groups.
Conclusion:
This study demonstrates the key dosimetric metrics and clinical variables. By applying Causal ML, we establish a framework that supports generating causally informed recommendations for radiotherapy planning and ORN management.