Artificial Neural Network-Based Variable Importance Analysis of Prognostic Factors Related to Radiation Pneumonitis in Patients with Lung Cancer: Preliminary Study

Abstract

Radiation pneumonitis (RP) is a major radiation-induced lung injury in patients with lung cancer. When an extremely high risk of severe RP is predicted, it is a critical issue that severely affects radiation dose and treatment planning. It is essential to assess which prognostic factor affects the occurrence of RP before initiating radiation treatment. In this study, we aimed to identify the variable importance (VI) of prognostic factors related to RP by using an artificial neural network (ANN) containing complex association between several prognostic factors. We reviewed 110 cases in patients with lung cancer who received radiation therapy (RT) from August 2000 to December 2018. The multi-layer perceptron algorithm, which is a back-propagation ANN algorithm, was implemented by SPSS Modeler (verl3.1, IBM SPSS Inc., Chicago, IL). The fifteen input variables were set, and the target variable was the occurrence of RP. The VI, which indicates the effect of each prognostic factor on the occurrence of RP, was analyzed by the variance-based method. To evaluate the VI of the ANN, we qualitatively compared the VI of the ANN with the odds ratios (OR) obtained from previously published literature. Patients who had an RP grade >2 comprised 13.6%. The accuracy of the ANN model was 76.92% and the and the area under the curve (AUC) was 0.774. From the results of the VI from the ANN, mean lung dose (MLD) had the highest VI among the prognostic factors at 15.96%. The OR of interstitial lung disease (ILD) was the highest at 25.70. While the VI of the ANN includes complex associations of the prognostic factors, ORs include independent associations between exposure and outcome. This preliminary study is meaningful as it proposes a method to quantify the VI of prognostic factors related to patient outcome using even simple ML algorithm. By further acquiring the more refined big data, optimizing the ML model, and performing clinical validation, the accurate VI can be adopted for personalized RT planning.