DeepDiagnosis: DNN-Based Diagnosis Prediction from Pediatric Big Healthcare Data

Abstract

Mining electronic health records (EHRs) has been considered as a major decision-making tool for clinical diagnosis. In fact, it is difficult to extract the valuable information from EHRs due to free-text writing, incomplete description, and high variabilities of diseases. Especially for pediatric EHRs, the shortage of experienced pediatricians as well as complex environmental factors such as seasonal variations, cross infections from kindergartens, make it extremely challenging to conduct a precise diagnosis. To address those challenges, we proposed DeepDiagnosis, a novel deep neural network-based diagnosis prediction algorithm by mining massive pediatric EHRs. First, we pre-process the unstructured EHRs dataset in Chinese and transfer them into sentence vectors by natural language processing technologies. Second, we construct the bidirectional recurrent neural networks (BiRNN) model to catch the patients' clinical symptoms as well as their interaction. Finally, we train and evaluate our model using a real-world dataset containing 81,476 pediatric EHRs. Experimental results show that the proposed method outperforms many baseline methods.