基于深度学习的人体低氧状态识别

通过低氧实验提出一种快速识别人体低氧状态的方法.通过搭建深层神经网络训练实验数据识别氧气体积分数(16%~21%)与人体可耐受极端低氧气体积分数(15.5%~16%)条件下光电容积脉搏波(photoplethysmography, PPG)信号, 获得人体生理状态的模式识别网络.经测试该网络的识别正确率可达92.8%.利用混淆矩阵及接受者操作性能(receiver operating characteristic, ROC)曲线分析, 混淆矩阵的训练集、验证集、测试集、全集识别正确率分别达到97.9%、94.8%、92.8%和96.3%, AUC (area under curve)值接近1, 认为该网络分类性能优良, 并且可在4 s内完成整个识别过程. 

Recognition of human hypoxic state based on deep learning

Due to the development of industrialization, low-oxygen environment has become common in the confined spaces of construction industries, chemical industries, military, urban underground spaces, and poorly ventilated crowed areas and caused a large number of hypoxic injuries. The traditional method of preventing hypoxic injuries is to monitor the oxygen concentration in the environment without considering the difference in oxygen tolerance limits when the human body is in different physiological states. Photoplethysmography (PPG) can comprehensively reflect physiological information, including heart rate, blood pressure, blood oxygen saturation, cardiovascular blood flow parameters, and respiratory rate. When the human body enters a hypoxic environment, the physiological parameters change rapidly, resulting in a change in the PPG signal. By measuring the PPG signal of the human body, the physiological state is considered to determine whether the human body reaches the oxygen tolerance limit. This study proposed a method for quickly identifying the hypoxic state of the human body using hypoxia experiment. According to the latest research on aviation medicine, mountain medicine and naval submarine medicine, 15.5% oxygen volume fraction can guarantee the basic life safety of personnel. Through the training experimental data of a constructed deep neural network, the PPG signal of a human in normal oxygen volume fraction (16% -21%) and extremely low-oxygen volume fraction (15.5% -16%) was determined to obtain the pattern recognition network of human physiological state. After testing, the recognition accuracy of the network could reach 92.8%. Using the confusion matrix and receiver operating characteristic curve analysis, the accuracy rate of training set, verification set, test set, and ensemble recognition of the confusion matrix reached 97.9%, 94.8%, 92.8%, and 96.3%, respectively. The area under the curve value is close to 1, the network classification performance is excellent, and the entire identification process could be completed within 4 s.