Room: Exhibit Hall
Purpose: We applied machine learning techniques to predict the occurrence of late GI toxicity among patients in a Phase I/II trial of SBRT for unresectable pancreatic cancer.
Methods: Dosimetric data for 15 patients who received 30-45Gy in 3 fractions were included (4 developed late Grade 3+ bowel toxicity). The tumor size, prescription dose, PTV V105%-V125%, and Duodenum D0.01-D5.0cc, V15-30Gy were selected as training features for machine learning using scikit-learn in Python. A feature selector was employed to rank the relative importances based on univariate statistics. Linear binary classifiers, i.e., k-nearest neighbour (kNN), logistic regression (LR) and support vector machine (SVM) models, were trained with a variety of leading features. Model hyper-parameters and selected number of features were cross validated by the area under the receiver operating characteristic curve (AUROC). The best model (greater AUROC and lower variance) was re-trained to derive a decision boundary where the severe toxicity probability was 0.5.
Results: The features with leading importances were PTV:V120%, Duodenum: V26Gy, GTV and PTV size, Duodenum: D0.1cc, followed by others. With the first three features, both LR and SVM models accurately predicted toxicity (meanÂ±standard deviation AUROC 1.0Â±0.0). The AUROC decreased to 0.95Â±0.5 when 5+ features were used. Using single neighbor with 10 features was found best for kNN but the model underperformed (AUROC 0.86Â±0.14). In a 3D feature space, the re-trained LR model constructed a planar decision boundary that clearly separated patients with observed events.
Conclusion: Clinical data analysis with machine learning provided useful insights on the risk of GI toxicity in this small cohort of patients. As a supplement to other clinical knowledge, the learned model is potentially useful for identifying patients with high risks based their tumor size and treatment plan dose distributions.