Abstract: How can we effectively carry out the feature learning of high-dimensional and heterogeneous medical electronic data to optimize the risk prediction of concurrent medical use in patients? To address the problem, this paper proposed a method of multi-stage deep feature learning. Firstly, we performed the feature learning of medical use data with temporal properties by combining deep learning models of long short-term memory (LSTM) and auto-encoder (AE), and generated the synthetic factor of concurrent medical use with bisecting k-means clustering method. Secondly, we constructed two types of feature vectors for patients to predict adverse event risk, and analyzed the associated factors of high risk. Finally, we compared the performance of the proposed method with existing methods on real-word dataset, and the results show that the proposed method increases the accuracy by 5%~10%, and reduces the false rate by 3%~5% in the risk prediction of concurrent medical use.

Key words: deep learning, feature learning, medical electronic data, concurrent medical use, adverse events