基于时空信息分段融合模型的术中低血压预测

Two-Stage Spatio-Temporal Information Fusion Model for Intraoperative Hypotension Prediction

准确提前预测术中低血压的发生对术中选择紧急方案以及降低患者术后的不良风险和死亡率具有积极作用。术中低血压预测目前主要基于术中患者的各项生理指标数据,但现有方法未能充分考虑多指标间的时间信息和空间信息。针对以上问题,提出基于时空信息分段融合模型的术中低血压预测,先使用全卷积网络或残差网络提取局部空间信息,再使用门控循环单元提取时间信息并进行预测。通过对比术中低血压预测常用的深度神经网络模型,在原始和填补的临床数据的术中低血压预测中,不仅提高了低血压事件的预测准确性,还在面对数据填补时表现出一定的容忍度,能够应对噪声和不确定性的影响。

Accurately predicting intraoperative hypotension in advance has a positive impact on the selection of emergency in-terventions during surgery and the reduction of adverse risks and mortality rates in postoperative patients.Currently,the prediction of intraoperative hypotension is mainly based on various physiological indicators of patients during surgery,and existing methods fail to adequately consider the temporal and spatial information among multiple indicators.To address these issues,a two-stage spa-tio-temporal information fusion model for intraoperative hypotension prediction is proposed.It first utilizes either a fully convolution-al network or a residual network to extract local spatial information,and then employs gated recurrent units to capture temporal infor-mation for prediction.By comparing with commonly used deep neural network models for intraoperative hypotension prediction,the proposed model not only improves the prediction accuracy of hypotensive events in both original and imputed clinical data but also exhibits a certain tolerance when dealing with data imputation,thus effectively addressing the impact of noise and uncertainty.