Augmenting Embedding with Domain Knowledge for Oral Disease Diagnosis Prediction

In this paper, we propose to add domain knowledge from the most comprehensive biomedical ontology SNOMED CT to facilitate the embedding of EMR symptoms and diagnoses for oral disease prediction. We first learn embeddings of SNOMED CT concepts by applying the TransE algorithm prevalent for representation learning of knowledge base. Secondly, the mapping from symptoms/diagnoses to biomedical concepts and the corresponding semantic relations defined in SNOMED CT are modeled mathematically. We design a neural network to train embeddings of EMR symptoms and diagnoses and ontological concepts in a coherent way, for the latter the TransE-learned vectors being used as initial values. The evaluation on real-world EMR datasets from Peking University School and Hospital Stomatology demonstrates the prediction performance improvement over embeddings solely based on EMRs. This study contributes as a first attempt to learn distributed representations of EMR symptoms and diagnoses under the constraint of embeddings of biomedical concepts from comprehensive clinical ontology. Incorporating domain knowledge can augment embedding as it reveals intrinsic correlation among symptoms and diagnoses that cannot be discovered by EMR data alone.