Prediction Of Intraoperative Hypotension Based On Multi-view Feature Enhancement And Fusion

Intraoperative hypotension is an independent risk factor for perioperative adverse outcomes and is associated with many postoperative complications.It is of clinical practice and theoretical significance to study the real-time prediction model of intraoperative hypotension based on heterogeneous multi-source patient data.Firstly,in view of the large sparsity of the characteristics of seven time series data,such as average arterial pressure heart rate,blood oxygen saturation,body temperature,pulse,end-of-breath carbon dioxide,etc,monitored intraoperatively,the interpolation method closest to the characteristics of medical time series was searched from the perspective of the reduction of medical features with diagnostic significance,so as to ensure the sensitivity and specificity of the prediction model.Secondly,a multi-view learning model based on feature fusion is constructed.From the perspective of as many physiological or pathological features as possible,multi-view feature extraction based on context perception is performed.For static features,in order to reduce the number of features,using light GBM combined with the doctor’s experience,from the 64 original static features,40 static features with high correlation and discrimination were selected to form a static feature view.At the same time,from the perspective of data dimension expansion,the staged time series statistical characteristics of each type of time series data are extracted to increase the accuracy of model prediction.According to this,the features of time series data are enhanced from two perspectives of time domain and space domain,so as to explore the long-term and shortterm dependence of time series data.Formed the time domain feature view and the spatial domain feature view of time series data.A static feature view is constructed by combining the obtained time-series statistical features with the selected static features.An intraoperative hypotension prediction model based on multi-view feature enhancement and fusion was developed.The clinical validity of the model was verified by 174,434 surgical cases in a hospital in Wuhan and 6388 surgical cases in the open data set Vital DB of the Department of Anesthesiology and Pain Medicine,Seoul National University School of Medicine,Korea.The sampling frequency of real intraoperative timing data retained by a hospital is 5 minutes,while the sampling frequency of the real intraoperative time series data in Vital DB is 2seconds.Vital DB is closer to the actual intraoperative situation.On Vital DB,the area under the receiver’s operating characteristic curve of the model reached 0.940,and the sensitivity and specificity values reached 0.876 and0.877 respectively,which basically met the actual industrial use requirements of the machine learning model.On the dataset of a hospital in Wuhan,the area under the operating characteristic curve of the model receiver after feature enhancement has reached 0.872,and the sensitivity and specificity values reached 0.778 and0.792,respectively.Considering that the sampling frequency of time series data can reach2 seconds in the actual operation,this model has high clinical practicability.