Abstract
Breast cancer (BC) is a leading cause of cancer death among women in United States. Previous studies have indicated that Black American women have disproportionately higher mortality than non-Hispanic White American women. Existing studies have demonstrated that artificial intelligence (AI) and machine learning (ML) especially transfer learning (TL) could address BC health disparities by transferring information learned from a majority group (e.g., White American women) to minority groups (e.g., Black American women). However, these studies have the following limitations: (1) the performance will decrease significantly as limited patient samples for training can be collected in clinical settings; and (2) most of existing studies only leverage single-omics data without exploring multi-omics integration. We recently presented a transfer learning method by integrating two multi-omics data for reducing cancer disparities. However, the integration model was not optimized, and its performance in reducing disparities was not robust. To address these concerns, we propose a weighted multi-modal transfer learning framework called MOTLAB designed to optimize the multi-omics integration equipped with data augmentation to systematically mitigate racial disparities in BC. Specifically, we first calculated patient-patient similarity using the Pearson Correlation Coefficient (PCC), which were used to construct a weighted integration of multi-omics data. Then, we performed a nested grid search method to optimize the weight combinations for each omics modality, which were subsequently used in multi-omics data integration to generate input data of the transfer learning model. In addition, to reduce the impact of data imbalance problems for our TL model, we leveraged a data augmentation named Synthetic Minority Oversampling Technique (SMOTE) for the minority groups to further boost performance of reducing health disparities. Results based on a dataset of 1085 female BC patient samples from The Cancer Genome Atlas (TCGA) database suggested that MOTLAB with optimized weighted integration of three omics data (including mRNA, miRNA and methylation) outperformed existing multi-omics transfer learning models. Moreover, MOTLAB achieved better performance than single-omics and two-omics-integration transfer learning models as well as conventional mixed models and independent models for BC health disparities mitigation. We anticipate that MOTLAB will serve as a new approach to reduce health disparities in BC diagnosis, prognosis, and treatment, and be extensible to mitigate health disparities for other types of cancer.