HKDP: A Hybrid Knowledge Graph Based Pediatric Disease Prediction System

摘要

In this paper, we present a clinically pediatric disease prediction system based on a new efficient hybrid knowledge graph. Firstly, we automatically extract a set of triples by modeling and analyzing 1454 clinically pediatric cases, building a weighted knowledge graph based Naive Bayes. Secondly, to extract new prediction opportunities from heterogeneous data sources, we model and analyze both classically professional pediatrics textbooks and clinical experiences of pediatric doctors respectively in order to derive prediction rules. Thirdly, we mix up those rules with the weighted knowledge graph we built to propose a new hybrid knowledge graph which can carry on both the Bayesian reasoning and the logic calculation at the same time. Fourthly, in term of that hybrid knowledge graph, we further design a new multi-label classifier based on the well-known Bayesian Ranking for the disease prediction. Finally, we implement such a hybrid knowledge graph based disease pediatric prediction system (HKDP) which uses the descriptions of the patients' symptoms as the inputs so as to return the predicted candidates of diseases for a child. In our experiments, the comparisons with classical prediction methods prove the validity and advantage of our system, especially guaranteeing good balance between the interpretability and precision of predictions in HKDP.