ESTRO 2024 - Abstract Book

S4521

Physics - Machine learning models and clinical applications

ESTRO 2024

Figure 1: Receiver Operating Characteristic curve showing the AUC for the multivariate Logistic Regression and ElasticNet in predicting xerostomia outcomes.

The Logistic Regression and ElasticNet models incorporated (i.e pooled) data from 238 patients. Feature selection from the ElasticNet identified several variables, including the mean dose to the ipsilateral parotid gland (p<0.01), age (p<0.01), V20% for both ipsilateral and contralateral parotid glands (p<0.001, p<0.01), D98Gy for the oral cavity (p<0.001), V40% for the ipsilateral parotid gland (p<0.001) and V70% for the ipsilateral submandibular gland (p<0.01). The p-values were obtained from a permute t-test. Hyperparameter tuning for the ElasticNet model determined the regularization strength (alpha) to be 0.2, with a mixing parameter for L1 and L2 penalties (L1 ratio) of 0.7. Subsequently, the Logistic Regression model yielded an area under the curve (AUC) of 0.6, while the ElasticNet model outperformed with an AUC of 0.8 (see Fig 1).

Promising results were obtained from the DL model providing a mean validation classification AUC value of 0.79 (range: 0.63 to 0.96 in cross validation).

Conclusion:

A multivariate model (ElasticNet) for late xerostomia was constructed and based on doses to ipsi- and contralateral parotid, submandibular gland, age, as well as the dose to the oral cavity. The ElasticNet model outperformed the traditional Logistic Regression in predicting late xerostomia in terms of AUC. Our analysis showed that several dose volumes and other features correlated with xerostomia in contrast to clinical norm that focus on the mean dose. This difference shows the complexity of late xerostomia modelling. Our study demonstrates that dose response for xerostomia have several significant factors showing that there are more additional dose criteria, revealing the many aspects of the problem.