Machine learning-enhanced normal tissue complication probability modeling for late sciatic nerve toxicity prediction
Yongqiang Li1,2,3, Ping Li4,2,3, Ruoxi Wang5
1Department of Medical Physics, Shanghai Proton and Heavy Ion Center, Fudan University Cancer Hospital, Shanghai 201321, People's Republic of China.
Physics in Medicine and Biology
|October 8, 2025
Summary
A new machine learning model predicts late sciatic nerve toxicity (LSNT) after carbon-ion radiotherapy (CIRT) for sacrococcygeal chordoma and rectal cancer patients. This model identifies critical dose-volume thresholds to balance tumor control and neuroprotection.
Area of Science:
- Radiation Oncology
- Medical Physics
- Machine Learning in Medicine
Background:
- Late sciatic nerve toxicity (LSNT) is a significant concern in carbon-ion radiotherapy (CIRT) for sacrococcygeal chordoma (SC) and locally recurrent rectal cancer (LRRC).
- Accurate prediction of LSNT is crucial for optimizing treatment planning and minimizing patient morbidity.
- Existing normal tissue complication probability (NTCP) models may require enhancement for improved predictive accuracy in CIRT settings.
Purpose of the Study:
- To develop and validate a machine learning-enhanced NTCP model for predicting LSNT in SC and LRRC patients treated with CIRT.
- To identify critical dose-volume thresholds associated with different grades of LSNT.
- To provide evidence-based recommendations for CIRT planning to balance tumor control and sciatic nerve protection.
Main Methods:
- A dual-modeling approach was employed, analyzing data from 106 CIRT-treated SC/LRRC patients.
- A hybrid framework integrated the Lyman-Kutcher-Burman model with generalized machine learning techniques.
- Radiation dosimetry parameters (EUD, TD50, n, m) and biological parameters were analyzed using univariate and multivariate regression; model performance was assessed using ROC analysis, sensitivity, and specificity.
Main Results:
- 16.9% of patients developed grade ⩾1 LSNT, with no grade ⩾4 toxicity observed.
- Machine learning-enhanced multivariate analysis identified critical dose-volume thresholds: V62 for G1, V64 for G2, and D3cc for G3 LSNT.
- Equivalent Uniform Dose (EUD) > 61.1 Gy was found to significantly elevate G1 LSNT risk.
Conclusions:
- The developed machine learning-enhanced NTCP model effectively predicts LSNT in CIRT patients.
- Recommended dose constraints for CIRT planning include V62 ⩽ 6.2%, V64 ⩽ 4.69%, and D3cc ⩽ 32.3 Gy.
- The study provides critical insights into individual radiation sensitivity and aids in optimizing neuroprotection during CIRT for SC and LRRC.
