建立并验证一种新型多模态生理参数用于评估胸科手术麻醉深度

A novel multi-modality fusion method for assessing the depth of anesthesia in thoracic surgery

目的建立并验证一种新型多模态生理参数评估胸科手术麻醉深度的准确性。方法选择2022年10月-2023年2月于上海交通大学医学院附属胸科医院择期接受胸科手术的46例患者,年龄18~80岁,ASA分级Ⅰ至Ⅲ级。所有患者均接受全凭静脉麻醉。采集患者的多模态生理信号,包括五导联心电图(ECG)、有创桡动脉血压(以P1表示)、脉搏血氧饱和度、外科脉搏容积指数(SPI),脑电监测采用一次性无创脑电熵指数传感器测定患者熵指数波形。应用icollect数据采集软件连接麻醉监护仪,4通道同时采集并记录麻醉开始前患者清醒状态至麻醉苏醒期间的ECG、脑电图[EEG(以Ent100表示)]、P1和脉搏波形(Pleth)。由2名经验丰富的麻醉科医师将患者麻醉深度分为清醒、浅麻醉和适宜麻醉3种状态,并作为金标准。采用深度神经网络的方法提取并融合4种模态的生理信号(ECG、Ent100、P1、Pleth)特征,得到最终的麻醉深度评估的决策模型。计算多模态生理参数评估麻醉深度的准确率。结果最终纳入39例接受全身麻醉胸科手术患者的术中监测数据。去除无效数据或伪差后,共分析了23000个完整数据集,其中清醒状态、浅麻醉状态和适宜麻醉状态共有300、600、3000个数据集用于测试。当使用基本相同的网络结构时,单独输入ECG、EEG、P1和Pleth 4种单模态的数据,网络不收敛,训练loss均为1.092,无法提取对麻醉深度评估有用的决策信息;而ECG、EEG、P1和Pleth集成的多模态生理参数的模型评估患者麻醉状态的准确率为73.7%。结论本研究提出了一种基于衍生的多模态生理参数监测不同麻醉深度的方法,可较准确地评估患者术中麻醉深度。

Objective To establish and verify a novel multi-modality fusion method in assessing the anesthesia depth in thoracic surgery.Methods A total of 46 patients,aged 18-80 years old,American Society of Anesthesiologists(ASA)physical statusⅠ-Ⅲ,and scheduled for thoracic surgery between October 2022 and February 2023 were enrolled in this study.All patients received total intravenous anesthesia.Multimodal physiological signals were collected including 5-lead electrocardiogram(ECG),invasive radial artery blood pressure(represented as P1),pulse oximetry,and surgical pleth index(SPI).EEG monitoring was performed by using a disposable non-invasive EEG entropy index sensor.I-collect software was used to record ECG,EEG(represented as Ent100),P1 and pulse waveform(Pleth)in four channels from awake state to recovery from anesthesia.Two experienced anesthesiologists classified the depth of anesthesia into three states:awakeness,light anesthesia and adequate anesthesia.Deep neural network was used to extract and fuse the features of physiological signals(ECG,Ent100,P1,and Pleth)so as to obtain the final model for anesthesia evaluation.The hybrid multi-modal physiological parameter and individual single,modal parameters were compared to verify the accuracy of anesthesia depth.Results The intraoperative monitoring parameters of 39 patients were included in the final analysis.After removing invalid data and artifacts,a total of 23000 complete datasets were analyzed.There were 300,600,and 3000 datasets for testing in the states of awakeness,light anesthesia and adequate anesthesia,respectively.When data from four modalities was imputed as individual inputs,the network failed to converge and extract decision,making information for evaluating the depth of anesthesia,and the train loss was all 1.092.Nonetheless,when the model incorporated all the four modalities,it provided a useful evaluation in the anesthesia depth,with an accuracy of 73.7%.Conclusion We proposed a derived multimodal model for measuring the depth of anesthesia in thoracic surgery.It may be able to achieve more accurate assessment of the depth of anesthesia.