(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Predicting early whether a patient is likely to respond to a treatment supports personalized healthcare. Studies have shown that multimodal biomarkers provide complimentary information to support such prediction. However, data incompleteness hinders the efficient integration of multimodal biomarkers, and affects the performance of multimodal biomarker-based treatment outcome prediction. To address this issue, we proposed a deep learning-based data synthesis method to generate missing data for model training and improve the integration efficiency of multimodal biomarkers and hence improve the multimodal biomarker-based outcome prediction performance.
Methods: To efficiently leverage the incomplete multi-modal biomarkers, a deep learning-based method is proposed to synthesize the missing modality data based on the complimentary modality. An example of applying the proposed method to a two-modality multimodal outcome prediction framework is employed to demonstrate the usage of the proposed method.
Results: Based on a modularized multi-modal biomarker-based prediction model, we investigated the efficiency of the proposed data generation method using the PET images and microRNA features (genomics) of 305 oropharyngeal cancer cases collected under an IRB protocol. Only sixty of the data contain both the PET images and genomics. The performance of different feature extraction and classification methods trained with/without synthetic data has been compared as well. For all the tested method, the model trained with both synthetic and real data yield higher prediction performance in terms of accuracy and AUC with 24% increase as the maximum. The experimental results demonstrate the efficiency of the proposed method.
Conclusion: The proposed data generation method can be applied to any outcome prediction models having issues of incomplete multi-modal biomarkers and permit leveraging incomplete biomarkers for improving outcome prediction performance. It has practical applications in advancing personalized treatment research especially in the case of insufficient observations due to incomplete data.
Funding Support, Disclosures, and Conflict of Interest: This research was supported in part by NIH awards R01CA233873 and R21CA223799.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
