Premium
WE‐C‐AUD B‐05: Predicting Radiation‐Induced Cardiac Perfusion Defects Using a Fusion Model
Author(s) -
Chen S,
Zhou S,
Hubbs J,
Wong T,
BorgesNeto S,
Yin F,
Marks L,
Das S
Publication year - 2008
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.2962692
Subject(s) - receiver operating characteristic , sensitivity (control systems) , mars exploration program , support vector machine , artificial neural network , artificial intelligence , pattern recognition (psychology) , nuclear medicine , multivariate adaptive regression splines , cross validation , multivariate statistics , mathematics , medicine , computer science , statistics , regression analysis , physics , bayesian multivariate linear regression , astronomy , electronic engineering , engineering
Purpose: To predict radiation‐induced cardiac perfusion defects using a fusion model that combines the results of four separate models: feed‐forward neural networks (NNET), self‐organizing maps (SOM), support vector machines (SVM), and multivariate adaptive regression splines (MARS). Method and Materials: The database comprised 111 patients with left‐sided breast treated with radiotherapy (56 diagnosed with cardiac perfusion defects post‐radiotherapy). The four independent models (NNET, SOM, SVM, and MARS) were constructed using a small number of independently selected features. The four models were then fused to a final model by averaging their patient predictions. Patient predictions were generated by testing the models using ten‐fold cross‐validation, wherein 1/10 th of the data were tested, in turn, using models built with the remaining 9/10 th of the data. To account for the variance in patient predictions caused by the effect of data splitting, 10‐fold cross validation was repeated 100 times with random data splitting. Results: For the fused model, the area under the Receiver Operating Characteristics (ROC) curve for cross‐validated testing was 0.890±0.012 (sensitivity = 80.6±1.7%, specificity = 80.2±1.7%). It was superior to the individual models (NNET: ROC = 0.764±0.015, sensitivity = 72.9±1.5%, specificity = 72.4±1.6%; SOM: ROC = 0.769±0.013, sensitivity = 73.0±1.4%, specificity = 72.2±1.5%; SVM: ROC = 0.900±0.048, sensitivity = 87.3±6.2%, specificity = 86.0±6.1%; MARS: ROC = 0.802±0.009, sensitivity = 76.1±1.1%, specificity = 75.6±1.1%) either in regard to higher predictive capability or lower variance. The fused model identified the following features as most important in predicting radiation‐induced perfusion defects: generalized equivalent uniform dose (EUD) with exponent a = 0.7, 1.0, and 3.6, and hypertension. Other features such as V46, V47, obesity, pack years, and chemotherapy played a less important role. Conclusion: The fused model provides promise for prospectively predicting radiation‐induced cardiac perfusion defects with high accuracy and confidence (low variance).