Breast Cancer Classification with Electronic Medical Records Using Hierarchical Attention Bidirectional Networks

Precise breast cancer classification can aid the clinicians in making decision of diagnosis and therapy. The existing works mainly utilized medical images for breast cancer classification, ignoring the value of Electronic Medical Records (EMR). In this study, we investigate the problem of breast cancer classification with EMR data and propose a new breast cancer classification model based on Hierarchical Attention Bidirectional Networks (HA-BiRNN). Our proposed model supports the integration of the features of the diagnosis reports with the features of the context information in EMR. Specially, our proposed model first uses three HA-BiRNNs to extract features from one type of report in a hierarchical way. In each HA-BiRNN, a hierarchical neural network structure, consisting in two encoder layers of BiRNN (Bidirectional Recurrent Neural Networks), mirrors the hierarchical structure of diagnosis reports, and a hierarchical attention mechanism, consisting of two level attentions, attends to important elements within diagnosis report with word-level attention and sentence-level attention. Secondly, to learn the feature of context information, our proposed model uses two additional BiRNNs. Finally, we add one more recurrent neural network layer for classifying breast cancer based on patients' features generated by combing the reports' vectors. We evaluate our method on the real world breast cancer clinical reports, and results show that our method achieves higher performance on breast cancer classification.