Predictive Modelling of Toxicity Resulting from Radiotherapy Treatments of Head and Neck Cancer

TL;DR

This study developed a predictive model for severe dysphagia caused by radiotherapy in head and neck cancer patients, but the current model shows limited predictive accuracy, highlighting the need for further improvement.

Contribution

The paper presents a novel approach using 3D dose distribution data and machine learning to predict radiotherapy toxicity in head and neck cancer patients.

Findings

Support vector classification achieved AUC of 0.54 on validation.

The model's predictive power is currently limited.

Work is ongoing to enhance feature engineering and modeling.

Abstract

In radiotherapy for head and neck cancer, the radiation dose delivered to the pharyngeal mucosa (mucosal lining of the throat) is thought to be a major contributing factor to dysphagia (swallowing dysfunction), the most commonly reported severe toxicity. There is a variation in the severity of dysphagia experienced by patients. Understanding the role of the dose distribution in dysphagia would allow improvements in the radiotherapy technique to be explored. The 3D dose distributions delivered to the pharyngeal mucosa of 249 patients treated as part of clinical trials were reconstructed. Pydicom was used to extract DICOM (digital imaging and communications in medicine) data (the standard file formats for medical imaging and radiotherapy data). NumPy and SciPy were used to manipulate the data to generate 3D maps of the dose distribution delivered to the pharyngeal mucosa and calculate metrics describing the dose distribution. Multivariate predictive modelling of severe dysphagia, including descriptions of the dose distribution and relevant clinical factors, was performed using Pandas and SciKit-Learn. Matplotlib and Mayavi were used for 2D and 3D data visualisation. A support vector classification model, with feature selection using randomised logistic regression, to predict radiation-induced severe dysphagia, was trained. When this model was independently validated, the area under the receiver operating characteristic curve was 0.54. The model has poor predictive power and work is ongoing to improve the model through alternative feature engineering and statistical modelling approaches.