Knowledge Learning Symbiosis for Developing Risk Prediction Models from Regional EHR Repositories

Secondary use of regional EHR data suffers several problems, including data selection bias and limited data size caused by data incompleteness. Here, we propose knowledge learning symbiosis (KLS) as a framework to incorporate domain knowledge to address the problems and make better secondary use of EHR data. Under the framework, we introduce three main categories of methods: knowledge injection to input features, objective functions, and output labels, where knowledge-enhanced neural network (KENN) was first introduced to inject knowledge into objective functions. A case study was conducted to build a cardiovascular disease risk prediction model on the type 2 diabetes patient cohort using regional EHR repositories. By incorporating a well-established knowledge risk model as domain knowledge under our KLS framework, we increased risk prediction performance both on small and biased data, where KENN showed the best performance among all methods.