基于FT-LSTM模型的临床事件诊断序列预测研究

针对不同患者的临床数据信息难以得到合理表示，且事件之间时间间隔不同，从而导致预测困难等问题，提出一种基于词向量表示，并添加时控单元的时控长短期记忆神经网络（FT-LSTM）预测模型。首先通过FastText方法对医学事件进行可解释性的向量表示，以更有效地捕获富含医学信息的概念关系；然后针对临床数据对时间戳有着强依赖性的现象，在原有LSTM模型的基础上设计时控门，以更好地捕获长短期信息，对事件信息进行建模，从而改善预测表现。在MIMICⅢ数据集上的实验结果表明，使用FT-LSTM模型预测的召回率、准确率皆高于多种对比模型，证明了该方法的有效性。

Research on intelligent diagnosis sequence prediction based on FT-LSTM model

Aiming at the problems of time series data extraction and medical concept expression difficulty in electronic medical record(EMR) field, this paper proposed a time-controlled long-term and short-term memory neural network(FT-LSTM) prediction model. The model efficiently captured multidimensional features, and further enabled accurate prediction of future clinical events. First, it used the FASTTEXT method to interpret vector representations of medical events, more effectively capture rich conceptual relationships in medical information. Then, it designed a time gate based on the original LSTM model to better capture the long-term and short-term information, model the event information to improve the prediction performance. The experimental results on the MIMICIII dataset show that, using the FT-LSTM model can predict event results with higher precision, which is of great significance.