Optimizing dosage in linear accelerator based on predictive analysis of radiation induced skin toxicity using machine learning techniques

Abstract

Modern radiotherapy planning and dose delivery utilize treatment planning systems (TPS) with medical linear accelerator (LINAC). A TPS is a specialized computer system that simulates values for the parameters of LINAC for radiation therapy delivery, including gantry movement, collimator angle, and dose calculation. The rule-based approach for a treatment plan is often found inaccurate due to the diverse nature of tumors and skin sensitivity in patients. Radiation-Induced Skin Toxicity (RIST) represents a significant adverse consequence of radiotherapy, impacting the well-being of cancer patients. To optimize radiation treatment and minimize RIST, the development of effective predictive and assessment methods is essential. Recent years have witnessed a surge in the application of artificial intelligence and machine learning to various aspects of radiation therapy, particularly in the prediction and grading of RIST. This study offers a comparative evaluation of the performance of diverse machine learning models for the screening and grading of RIST severity. Furthermore, it focuses on the identification of the most influential feature set for this classification task. Our dataset comprises 2000 records, incorporating 18 clinical attributes from patients who underwent radiotherapy treatment at an Indian hospital. We conduct thorough statistical analyses of features derived from these clinical attributes, utilizing different feature subsets for model training and evaluation. Intriguingly, we find that by utilizing just five selected clinical attributes, machine learning models achieve nearly equivalent performance to using the entire attribute set. In terms of model performance, Support Vector Machine excels in the screening task, while Multi-layer Perceptron stands out in the gradation task among single models. Within the ensemble methods, Random Forest surpasses AdaBoost in both the screening and gradation tasks.