Enhanced prediction of radiation-induced skin toxicity in breast cancer patients using a hybrid dosiomics-clinical model

Abstract

Objectives

This study aims to develop a predictive model for radiation-induced skin toxicity (RIST) in breast cancer patients using dosiomics features extracted from the dose distribution map within the clinical target volume (CTV).

Materials and methods

This study included breast cancer patients treated with 3D conformal radiation therapy (3D-CRT). Patients were categorized into low-grade (G0-G1) and high-grade (G2-G3) toxicity groups. Dosiomics features of CTV, clinical data of medical records, and dosimetric parameters of dose maps were extracted. Three predictive models were developed: a dosiomics model using CTV-based features, a hybrid dosiomics-clinical model (HDO), and a hybrid dose-volume histogram-clinical model (HDV). Machine learning algorithms (support vector machines and random forests) were used to build the models and their performances were assessed using the area under the receiver operating characteristic curve (AUC).

Results

Thirty-two patients (42%) experienced high-grade RIST (CTCAE grade ≥2) following breast radiation therapy (RT). The HDO model demonstrated superior predictive performance, attaining an AUC of 0.78, significantly higher than the HDV and single predictive models. In the dosiomics-based features group, the major axis length from the shape class is one of the most relevant features for skin toxicity (grade ≥2). In the clinical parameters group, chemo regimen, receptor state, and hormonal treatment showed significant correlation with skin toxicity (p-value<0.05). In the DVH factors group V105 cc, V 110%, V107%, and Breast CTV revealed a significant correlation with skin toxicity.

Conclusion

The developed predictive model utilizing dosiomics features demonstrated superior performance compared to dose volume histogram (DVH) based methods, with an AUC of 0.78, leading to early prediction of skin toxicity among breast cancer patients who had received RT. Moreover, our results suggest that the integration of dosiomics features with clinical parameters significantly improves the predictive power of models.