Combining Reverse Temporal Attention Mechanism and Dynamic Similarity Analysis for Disease Prediction

In recent years, the explosive growth of the Internet has made it an indispensable part of our daily study, work and life. With the rapid development of the medical industry and information technology, electronic health record (EHR) system has been widely used, and more and more researchers choose to conduct experiments based on EHR data sets. However, due to the complexity and heterogeneity of patient EHR data, and electronic health record is a long-term process. Therefore, it is difficult for researchers to analyze EHR data sets for experiments such as predictions and recommendations. In this paper, we propose a disease prediction model (RTAMDSA model), which integrates Reverse Time Attention Mechanism (RTAM) and Dynamic Similarity Analysis (DSA). The model mainly includes a patient representation module, a similarity analysis module and a disease prediction module. The patient representation module uses a reverse temporal attention mechanism with two levels of attention, to run Recurrent Neural Network in reverse chronological order, to obtain the embedding representation of the patient from the patient's historical disease trajectory, so that the patient's most recent visit records receive more attention than the old records, because the patient's recent visit information is more informative. The similarity analysis module performs searches using dynamic similarity analysis, to calculate the similarity of each patient for each visit, and obtain the specific patient sequences that are most similar to the target patient. The disease prediction module integrates the obtained patient embeddings and similar patient sequences to output the final prediction result for the target patient. Comparing with the baseline models on the MIMIC-III dataset, experimental results show that the RTAMDSA model improves prediction accuracy on the disease prediction task, exhibits better robustness and stability, and the reverse time attention mechanism provides good interpretability for the model.