Radiotherapy Induced Xerostomia Prediction Through Cluster Models Incorporating 3D Spatial Dose Within the Parotid Gland and Machine Learning Techniques
M Chao1*, I. El Naqa2, R. Bakst1, Y. Lo1, J. A. Penagaricano2 (1) The Mount Sinai Medical Center, New York, NY, (2) Moffitt Cancer Center, Tampa, FL
Presentations
MO-C930-IePD-F5-3 (Monday, 7/11/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 5
Purpose: A cluster model incorporating inhomogeneous dose distribution within the parotid gland was developed and validated retrospectively for radiotherapy (RT) induced xerostomia prediction with machine learning (ML) techniques.
Methods: 155 patients with head-and-neck cancer (HNC) were analyzed in the study. For each enrolled patient, classified with binary xerostomia levels, sixty clusters were obtained with threshold doses from 1Gy to 60Gy at 1Gy step size. Neighborhood component analysis (NCA) was adopted to select cluster features, which were subsequently fed into four supervised ML models for xerostomia prediction: (1) support vector machines (SVM), (2) k-nearest neighbor (kNN), (3) naïve Bayes (NB), and (4) random forest (RF). The performance of each model was evaluated using cross validation resampling with the area-under-the-curves (AUC) of the receiver-operating-characteristic and compared to that of the mean dose based logistic regression model. Testing with independent dataset was assessed with accuracy, sensitivity, and specificity for these models and three cluster connectivity choices.
Results: Feature clusters selected by NCA peaked in three threshold dose ranges: 5-15Gy, 25-35Gy, and 45-50Gy. Mean dose predictive power was 15% lower than that of cluster based models from the AUC comparison. Model validation demonstrated that kNN model outperformed other three models but no substantial difference was observed. Applying the fine-tuned models to testing data yielded that the accuracy obtained from SVM, kNN and NB models was better than that of RF while SVM model showed the best sensitivity and kNN model delivered consistent sensitivity and specificity. This agrees with cross validation. Minimal different predictions were observed with three connectivity choices.
Conclusion: The proposed cluster model, in contrast to mean dose, has shown its improvement as the xerostomia predictor. When combining with ML techniques, it could provide a clinically useful tool for xerostomia prediction and facilitate decision making during radiotherapy planning for patients with HNC.
Keywords
Taxonomy
Contact Email