(Monday, 7/15/2019) 4:30 PM - 6:00 PM
Room: Stars at Night Ballroom 2-3
Purpose: Accurately predicting treatment outcome can provide a personalized treatment plan and follow-up schedule. In this study, we aim to develop a reliable and automatic multi-objective ensemble deep learning (MoEDL) model to predict patients at high risk of treatment failure after radiotherapy in lung cancer using features extracted from electronic health records (EHRs).
Methods: The dataset used in this study contains EHRs of 801 patients never being disease-free and 244 disease-free patients after lung cancer radiation therapy. The proposed MoEDL consists of three phases: 1) dynamic ensemble training; 2) multi-objective model selection; and 3) evidential reasoning (ER) fusion based prediction. In the training phase, we employ a number of different deep perceptron networks whose structures are dynamically changed to handle heterogeneity, sparseness, incompleteness and random errors of EHRs data. A Snapshot Ensembles (SE) restart strategy is employed to obtain multiple candidate networks with no additional training costs. A model selection strategy based on multi-objective optimization is designed to obtain Pareto-optimal networks with balanced sensitivity, specificity and area under the curve (AUC). In the testing stage, each sample is fed into the selected Pareto-optimal networks and the final predictive outcome is obtained by fusing the probabilities of the selected network models using the ER approach.
Results: Five-fold cross-validation was performed in this study. Sensitivity, specificity, accuracy, and AUC for MoEDL are 0.77, 0.60, 0.74, and 0.72, respectively. MoEDL outperforms traditional deep perceptron network (DNN), support vector machines (SVM), and an improved group based multi-objective model (GMO).
Conclusion: A reliable and automatic MoEDL model by using low-dimensional clinical record features was proposed to predict treatment outcome after radiotherapy for lung cancer patients. The experimental results demonstrated that MoEDL can obtain better performance compared with other conventional methods.
Funding Support, Disclosures, and Conflict of Interest: The Fundamental Research Funds for the Central University under Grant JB181701.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
