Skip to Main Content
The amount of delivered radiation dose to the tumor in non-small cell lung cancer (NSCLC) patients is limited by the negative side effects on normal tissues. The most dose-limiting factor in radiotherapy is the radiation-induced lung toxicity (RILT). RILT is generally measured semi-quantitatively, by a dyspnea, or shortness-of-breath, score. In general, about 20-30% of patients develop RILT several months after treatment, and in about 70% of the patients the delivered dose is insufficient to control the tumor growth. Ideally, if the RILT score would be known in advance, then the dose treatment plan for the low-toxicity-risk patients could be adjusted so that higher dose is delivered to the tumor to better control it. A number of possible predictors of RILT have been proposed in the literature, including dose-related and clinical/demographic patient characteristics available prior to radiotherapy. In addition, the use of imaging features -- which are noninvasive in nature - has been gaining momentum. Thus, anatomic as well as functional/metabolic information from CT and PET scanner images respectively are used in daily clinical practice, which provide further information about the status of a patient. In this study we assessed whether machine learning techniques can successfully be applied to predict post-radiation lung damage, proxied by dyspnea score, based on clinical, dose-related (dosimetric) and image features. Our dataset included 78 NSCLC patients. The patients were divided into two groups: no-deterioration-of-dyspnea, and deterioration-of-dyspnea patients. Several machine-learning binary classifiers were applied to discriminate the two groups. The results, evaluated using the area under the ROC curve in a cross-validation procedure, are highly promising. This outcome could open the possibility to deliver better, individualized dose-treatment plans for lung cancer patients and help the overall clinical decision making (treatment) process.