Explainable Artificial Intelligence to Identify Dosimetric Predictors of Toxicity in Patients with Locally Advanced Non-Small Cell Lung Cancer: A Secondary Analysis of RTOG 0617

Purpose Dosimetric predictors of toxicity in patients treated with definitive chemoradiation for locally advanced non-small cell lung cancer are often identified through trial and error. This study used machine learning (ML) and explainable artificial intelligence to empirically characterize dosimetric predictors of toxicity in patients treated as part of a prospective clinical trial. Methods and Materials A secondary analysis of the Radiation Therapy Oncology Group (RTOG) 0617 trial was performed. Multiple ML models were trained to predict grade ≥3 pulmonary, cardiac, and esophageal toxicities using clinical and dosimetric features. Model performance was evaluated using the area under the curve (AUC). The best performing model for each toxicity was explained using the Shapley Additive Explanation (SHAP) framework; SHAP values were used to identify relevant dosimetric thresholds and were converted to odds ratios (ORs) with confidence intervals (CIs) generated using bootstrapping to obtain quantitative measures of risk. Thresholds were validated using logistic regression. Results The best-performing models for pulmonary, cardiac, and esophageal toxicities, outperforming logistic regression, were extreme gradient boosting (AUC, 0.739), random forest (AUC, 0.706), and naive Bayes (AUC, 0.721), respectively. For pulmonary toxicity, thresholds of a mean dose >18 Gy (OR, 2.467; 95% CI, 1.049-5.800; P = .038) and lung volume receiving ≥20 Gy (V20) > 37% (OR, 2.722; 95% CI, 1.034-7.163; P = .043) were identified. For esophageal toxicity, thresholds of a mean dose >34 Gy (OR, 4.006; 95% CI, 2.183-7.354; P < .001) and V20 > 37% (OR, 3.725; 95% CI, 1.308-10.603; P = .014) were identified. No significant thresholds were identified for cardiac toxicity. Conclusions In this data set, ML approaches validated known dosimetric thresholds and outperformed logistic regression at predicting toxicity. Furthermore, using explainable artificial intelligence, clinically useful dosimetric thresholds might be identified and subsequently externally validated. Dosimetric predictors of toxicity in patients treated with definitive chemoradiation for locally advanced non-small cell lung cancer are often identified through trial and error. This study used machine learning (ML) and explainable artificial intelligence to empirically characterize dosimetric predictors of toxicity in patients treated as part of a prospective clinical trial. A secondary analysis of the Radiation Therapy Oncology Group (RTOG) 0617 trial was performed. Multiple ML models were trained to predict grade ≥3 pulmonary, cardiac, and esophageal toxicities using clinical and dosimetric features. Model performance was evaluated using the area under the curve (AUC). The best performing model for each toxicity was explained using the Shapley Additive Explanation (SHAP) framework; SHAP values were used to identify relevant dosimetric thresholds and were converted to odds ratios (ORs) with confidence intervals (CIs) generated using bootstrapping to obtain quantitative measures of risk. Thresholds were validated using logistic regression. The best-performing models for pulmonary, cardiac, and esophageal toxicities, outperforming logistic regression, were extreme gradient boosting (AUC, 0.739), random forest (AUC, 0.706), and naive Bayes (AUC, 0.721), respectively. For pulmonary toxicity, thresholds of a mean dose >18 Gy (OR, 2.467; 95% CI, 1.049-5.800; P = .038) and lung volume receiving ≥20 Gy (V20) > 37% (OR, 2.722; 95% CI, 1.034-7.163; P = .043) were identified. For esophageal toxicity, thresholds of a mean dose >34 Gy (OR, 4.006; 95% CI, 2.183-7.354; P < .001) and V20 > 37% (OR, 3.725; 95% CI, 1.308-10.603; P = .014) were identified. No significant thresholds were identified for cardiac toxicity. In this data set, ML approaches validated known dosimetric thresholds and outperformed logistic regression at predicting toxicity. Furthermore, using explainable artificial intelligence, clinically useful dosimetric thresholds might be identified and subsequently externally validated.