MCPL-Based FT-LSTM: Medical Representation Learning-Based Clinical Prediction Model for Time Series Events

Abstract

Large collections of electronic medical records (EMRs) provide us with a vast source of information on medical practice. However, the utilization of these data to support clinical decisions is still limited. Extracting useful patterns from such data is particularly challenging because the data are variable longitudinal, sparse, and heterogeneous. Therefore, in this paper, we propose the MCPL-based FT-LSTM, a clinical event prediction method based on medical concept representation learning. On one hand, inspired by FASTTEXT, we have developed an interpretative vector representation of medical events in EMRs, which enables us to capture the medical concept information effectively so that the patient's clinical data can be represented more reasonably. On the other hand, we propose a novel time-controlled long short-term memory (LSTM) prediction model, which adds time-control units to the original LSTM model. The model can describe the variable time intervals in EMRs, better capture long-term, and short-term information, and eliminate the strong dependence of clinical data on timestamps; thus, improving the model's prediction performance for clinical events. Through extensive experiments on the MIMICIII dataset, we demonstrate that the MCPL-based FT-LSTM achieves higher precision in the field of clinical event prediction, which is of great significance for the medical information research.