Health Causal Probability Knowledge Graph
Hong Qing Yu · 2020
Currently, most of the health-data research concentrates on applying Deep Learning technologies for prediction and reasoning. Deep Learning processes build the prediction model purely based on fitting weights on the raw data inside multiple neural layers, which is difficult to explain the prediction outputs. However, telling ‘WHY’ is crucial for healthcare research. The major difficulty to explain in Deep Learning models is a lack of knowledge-based analysis environment that not only can model the knowledge in a machine-understandable way but also can create causal probability relations inside the knowledge. In our research, we propose a Causal Probability Description Logic (CPDL) framework that extended the current Description Logic (DL). The key extension is to have a two-layer DL representation. One layer represents causality knowledge. The other layer takes observation inputs e.g. symptoms for generating a runtime probability knowledge graph based on the previous layer's knowledge. The CPDL framework can support probability-based causal reasoning tasks in a transparent and human-understandable way. CPDL can be easily implemented using existing programming standards such as OWL, RDF, SPARQL and probability network programming libraries. The experimental evaluations extract 383 common disease conditions from the UK NHS (National Healthcare Service) and enable automatically linked 418 condition terms from the DBpedia dataset. The CPDL-based knowledge graph can support disease prediction with traceable pieces of evidence behind the ranking results.