摘要目的:构建基于12导联体表心电图（ECG）图像对患者3、6 h内动脉血乳酸水平进行预测的深度学习模型（1种人工智能算法），探究其辅助评估患者微循环水平的效能。方法:本研究为回顾性、单中心研究。收集2015年1月9日至2021年12月30日就诊于上海交通大学医学院附属瑞金医院进行动脉血乳酸检测且6 h内有心电图记录的2 350例患者的年龄、性别和心电图图像，分别构建La1-ECG（采集心电图与动脉血乳酸标本时间差±3 h，包含心电图1 862份）、La2-ECG（采集心电图与动脉血乳酸的标本时间差±6 h，包含心电图2 350份）2个心电图集。以患者动脉血乳酸水平为分类标准（正常：≤1.5 mmol/L；轻度异常：1.5~5.0 mmol/L；重度异常：>5.0 mmol/L）构建深度学习模型并进行训练、验证及测试。对比算法在2个心电图集的精确率、召回率、受试者工作特征曲线下面积和F1指数。结果:共纳入2 350例患者，年龄（68.39±18.12）岁，其中男1 372例（58.38%，1 372/2 350）。基于体表心电图的图像建立的深度学习模型在La1-ECG中对于乳酸值的多分类评估的精确率、召回率、受试者工作特征曲线下面积（ AUC）和F1指数分别为80.27%、79.93%、0.94、0.80。在La2-ECG中对于乳酸值的多分类评估的精确率、召回率、 AUC和F1指数分别为78.29%、78.55%、0.92、0.78。 结论:基于体表12导联心电图图像的深度学习模型可以较好地分层预测动脉血乳酸水平，辅助临床医师进行患者的微循环情况评估。

abstractsObjective:To establish an artificial intelligence model based on 12-lead body surface electrocardiogram (ECG) images using deep learning algorithm, and to evaluate and diagnose the arterial blood lactate level of patients noninvasively and rapidly.Methods:It was a retrospective single-center study. ECG images were collected from 2 350 patients who underwent both arterial blood lactate testing and ECG within 6 hours from January 9, 2015 to December 30, 2021, in Ruijin Hospital, Shanghai Jiao Tong University School of Medicine. Two ECG sets [La1-ECG (arterial blood lactate testing and ECG within 3 h, 1 862 ECG recordings ) and La2-ECG (arterial blood lactate testing and ECG within 6 h, 2 350 ECG recordings) ] were constructed. Three different arterial blood lactate levels (normal: ≤1.5 mmol/L; mild abnormal: 1.5-5.0 mmol /L; severe abnormal: >5.0 mmol/L) were the standard to train, verify and test an artificial intelligence model based on deep learning. The precision, recall, area under the receiver operating characteristic curve ( AUC) and F1 score of the algorithm in the two ECG sets were compared. Results:A total of 2 350 patients were included in this study, with an average age of (68.39±18.12) years old and 1 372 males (58.38%, 1 372/2 350) . The precision, recall, AUC and F1 score of the deep learning model based on the 12-lead body surface ECG images for the multi-class evaluation of lactate values in La1-ECG were 80.27%, 79.93%, 0.94, and 0.80, respectively. The precision, recall, AUC and F1 score for the multi-class evaluation of lactate in La2-ECG were 78.29%, 78.55%, 0.92 and 0.78, respectively. Conclusion:The deep learning algorithm based on the 12-lead ECG image of the body surface can predict the arterial blood lactate level well, which may assist clinicians to evaluate the microcirculation of patients.