Multi-factor Prediction Model Of Lung Injury In Lung Cancer Patients With Radiation Therapy Based On Machine Learning Algorithm

Purpose:The present study aims to investigate the factors associated with radiation-induced lung injury（RILI）and develop a predictive model in lung cancer patients who receive conventional intensity-modulated radiation therapy（IMRT）.The ultimate goal is to establish a more precise RILI prediction model to guide clinical radiation physicists in designing more accurate radiation therapy plans for lung cancer patients,thereby improving the efficacy of radiation therapy and reducing the incidence of RILI after radiation therapy.Ultimately,this will lead to an improvement in the quality of life for post-treatment lung cancer patients.Methods:We collected clinical data,tumor-related information,and dose-related information of 216 lung cancer patients who underwent routine IMRT treatment in the radiotherapy department of the First Affiliated Hospital of Jinzhou Medical University from June 2019 to September 2022,including gender,age,smoking history,tumor size,location,and dose-related parameters such as V5 of the whole lung and ipsilateral lung and mean lung dose.We used SPSS29.0 software to perform a univariate analysis to explore the relationship between these parameters and radiation pneumonitis and identify the factors associated with radiation pneumonitis.We included univariate data with statistical significance in the multi-factor joint prediction model.We constructed multiple multi-factor joint prediction models with different input variable combinations using the Ridge Regression algorithm,Extreme Gradient Boosting algorithm,and XGBoost+MLP joint algorithm model based on computer machine learning methods.In the verification set,area AUC value under receiver operating characteristic curve,accuracy rate,precision,recall rate,F1-score and other evaluation indicatorswere used to evaluate the prediction efficiency of each model.Result:Univariate analysis showed that the occurrence of grade≥2 radiation pneumonitis was not significantly associated with gender,age,smoking history,tumor histology,radiotherapy and chemotherapy modes,tumor location,whole lung volume,or ipsilateral lung volume（P>0.05）.However,there was a significant correlation between the occurrence of radiation pneumonitis and MLD,V5,V10,V15,V20,V30,V40,as well as tumor-related parameters such as PCTV volume,V_PCTV/V _{whole lung},and V_PCTV/V_{ipsilateral lung}（P<0.05）.Based on the ridge regression algorithm,four multiple-factor prediction models（Model A,B,C,and D）were constructed,and the validation results in the validation set were as follows:Model A（AUC:0.7658,accuracy:0.7727,precision:0.5455,recall:0.5455,F1-score:0.5455）,Model B（AUC:0.7521,accuracy:0.7727,precision:0.5565,recall:0.4545,F1-score:0.5000）,Model C（AUC:0.7493,accuracy:0.7727,precision:0.5556,recall:0.4545,F1-score:0.5000）,and Model D（AUC:0.7493,accuracy:0.7500,precision:0.5000,recall:0.3636,F1-score:0.4211）.Additionally,four multiple-factor prediction models（Model A,B,C,and D）were constructed based on the extreme gradient boosting（XGBoost）algorithm,and the validation results in the validation set were as follows:Model A（AUC:0.7355,accuracy:0.8182,precision:0.6667,recall:0.5455,F1-score:0.6000）,Model B（AUC:0.8430,accuracy:0.7955,precision:0.5625,recall:0.8182,F1-score:0.6667）,Model C（AUC:0.8347,accuracy:0.7727,precision:0.5263,recall:0.9091,F1-score:0.6667）,and Model D（AUC:0.8499,accuracy:0.8182,precision:0.6000,recall:0.8182,F1-score:0.6923）.Finally,a prediction model（Model E）was constructed based on the XGBoost+MLP combined algorithm,and the validation results in the validation set were as follows:AUC:0.8892,accuracy:0.8636,precision:0.6471,recall:1.0000,F1-score:0.7857.Conclusion:The XGBoost+MLP combined algorithm,which is based on computer machine learning and deep learning algorithms,has been used to construct a multifactor joint prediction model for radiation-induced lung injury（RILI）.This model exhibits excellent predictive performance,and can provide clinical guidance for healthcare workers to predict the occurrence of≥2 grade RILI in patients caused by new radiotherapy plans.