关键词: 脑血管疾病, 重症监护病房（ICU）, 双向长短时记忆, 多通道, 注意力机制, 死亡风险预测

Abstract: Death risk prediction refers to the prediction of the risk of death for a period of time in the future based on the monitoring data of the patient’s clinical signs. For ICU patients, it is possible to make targeted clinical diagnosis for patients through the prediction of death risk, and rationally arrange limited medical resources. Based on the clinically used MEWS and Glasgow coma score scale, aiming at the 17 physiological parameters monitored by ICU patients, a multi-channel ICU cerebrovascular disease death risk prediction model is proposed. First, the concept of multi-channel is introduced and applied to the BiLSTM model to highlight the role of each physiological parameter in predicting the risk of death. Second, the Attention mechanism is used to improve the prediction accuracy of the model. The experimental data comes from the MIMIC [Ⅲ] database, from which 16?260 records of 3?080 patients with cerebrovascular diseases are extracted for this study. In addition to six sets of hyperparameter experiments, the proposed model is compared and analyzed with LSTM, Multichannel-BiLSTM, and logistic regression and support vector machine four models. Accuracy, Sensitive, Specificity, AUC-ROC and AUC-PRC are used as evaluation indicators. The experimental results show that the performance of the proposed model is better than other models, and the AUC value reaches 94.3%.

Key words: cerebrovascular disease, intensive care unit（ICU）, bi-directional long short-term memory（BiLSTM）, multichannel, attention mechanism, death risk prediction