B1953
Title: Targeted learning to predict toxicity impact on survival in advanced lung cancer patients
Authors: Gilmer Valdes - UCSF (United States) [presenting]
Abstract: The aim is to predict the likelihood and impact of grade 2 pulmonary toxicities on survival among lung cancer patients receiving proton radiation therapy (PBT) using machine learning techniques. Data from 965 patients across 17 institutions were analyzed for grade 2 toxicities. We employed a double 10-fold cross-validation technique for hyperparameter tuning in Gradient Boosting and Lasso algorithms. Balanced Accuracy (BA) and Area Under the Curve (AUC) metrics were used to assess model performance. Targeted learning analyzed the toxicities' causal effect on survival. Of the patients, 256 (28.2\%) had grade 2 toxicities. Key variables included technique used, concurrent chemotherapy, and total radiation dose. Centers using pencil beam scanning (PBS) had a lower toxicity rate (0.08) compared to older techniques (0.34). Abdominal compression also reduced toxicity. A model combining demographic and dosimetric variables achieved an AUC of 0.75 and BA of 0.67. Gradient Boosting outperformed other algorithms. Targeted learning revealed a 1\% decrease in 5-year survival for each percent increase in the likelihood of high-grade toxicities. In short, advanced machine learning identifies that using PBS, abdominal compression, and dose reduction to the normal lung can decrease the risk of grade 2 pneumonitis or dyspnea. These toxicities also adversely affect survival.