基于时序模型的麻醉期人体生命体征预测

Human vital signs prediction during anesthesia maintenance period based on time series model

针对现有麻醉维持期的生命体征预测方法存在个性化不足、精度较低和预测速度慢等问题,引入时序预测方法,提出一种基于保留机制的时序混合保留网络(TFR)。首先,使用变量选择网络和静态变量编码器网络对病人的生命体征进行编码,为网络提供病人的个性化信息;其次,通过保留编解码器网络对变量选择网络选择过的变量进行特征提取和编码;其次,使用多头注意力层融合过去和未来的特征;最后,使用输出网络进行多分位输出。在Vital DB和华西医院的数据集上分别进行训练和测试,实验结果表明,相较于基线模型TFT(Temporal Fusion Transformer):TFR在Vital DB数据集上的均方误差(MSE)和平均绝对误差(MAE)均没有增加,并节省了41.2%的推理时间;在华西医院的数据集上预测精度没有下降,并节省了32.1%的推理时间;在长期预测中,TFR也取得了更好的效果。可以看出,TFR预测更精确、更快速。

In view of the problems of insufficient personalization,low accuracy and slow prediction speed of existing vital sign prediction methods during anesthesia maintenance period,the time series prediction method was introduced,and a Temporal Fusion Retention network based on the retention mechanism(TFR)was proposed.Firstly,a variable selection network and a static variable encoding network were used to encode the patient’s vital signs,providing the network with personalized information about the patient.Then,the variables selected by the variable selection network were performed feature extraction and encoding through the retention encoder-decoder network.Afterwards,a multi-head attention layer was used to fuse past and future features.Finally,the output network was used for multi-quantile output.Training and testing were conducted on the Vital DB and West China Hospital datasets.Experimental results show that compared with baseline model TFT(Temporal Fusion Transformer),TFR has no increase in Mean Square Error(MSE)and Mean Absolute Error(MAE)on the Vital DB dataset,and saves 41.2%of inference time;it has no decrease in prediction accuracy on the West China Hospital dataset,and saves 32.1%of inference time;it also achieves better results in long-term prediction.It can be seen that TFR can make more accurate and faster predictions.