Integrating Empirical Knowledge into Multi-View Feature Attention Network for Disease Diagnosis

Anonymous

Abstract: As one of the currently significant problems in AI-enabled healthcare research, disease diagnosis based on the medical text has made substantial progress. However, the length of the diagnostic evidences is different, leading to the difficulty of capturing multi-scale features of each disease. And recent studies have discovered that structural knowledge from medical text is critical for disease diagnosis. This paper proposes integrating empirical knowledge of disease into a multi-view feature attention network to address these issues. The multi-view feature attention network employs multi encoders to capture segment information of diagnostic evidences of each illness. Meanwhile, we used an abductive causal graph constructed from medical text to extract the empirical knowledge representation of diseases by graph convolutional network. The evaluation conducted on the MIMIC-III-50 dataset and Chinese dataset demonstrates that the proposed method outperforms the structural knowledge-based state-of-the-art models.

Paper Type: long