Abstract: Abstract: With the emergence and wide application of Electronic Health Record （EHR）， the prediction model based on EHR data can be used for early detection and intervention of diseases. Heterogeneous attributes are ubiquitous in EHR data， but it is difficult to thoroughly exploit their information. Therefore， the method of heterogeneous attribute fusion provides an informative data representation basis for subsequent model training. This paper designs an efficient two-stage prediction model for solving the problems of time and cost in predicting critical illness. In the first stage of the model， coarse-grained prediction is performed on patient samples. Patients with low severity are initially screened out， which plays a key role in patient diversion. The second stage makes more fine-grained predictions of potentially critical patients based on the coarse filtering results of the first stage. The experimental verified that， after heterogeneous attribute fusion， when we select the first 6 time points to construct a non-temporal model， the two-stage model has better performance in both initial disease screening and disease prediction.

Key words: Key words: heterogeneous attribute fusion, disease screening, two-stage model