Intensive Care Unit False Arrhythmia Alarm Classification Based On Convolution Neural Network

With the development of science and information technology,monitoring equipment in the Intensive Care Unit(ICU)have gradually increased.In order to ensure the safety of patients,most of these devices are designed for high sensitivity,and often an alarm occurs when any indicator exceeds a threshold.However,a large number of studies have shown that more than 70% of the alarms generated in the ICU are false or nonactionable.While these alarms cause noise interference,it is even more dangerous that they may cause medical staff to be insensitive to the alarm,completely contrary to the original intention of the high-sensitivity design.Therefore,it is necessary to reduce the false alarm rate in ICU.In order to reduce the false alarm rate of ICU arrhythmia,we uses the ICU arrhythmia alarm dataset provided by the 2015 Physionet Challenge as training data,and designs a classification model based on convolutional neural network,which can directly deal with time series and avoid manually extracting features.In the model design,in order to extract features from multi-modal time series signals,we combines the grouping convolution strategy to construct two basic network structures,i.e.deep group convolutional neural network(DGCN)and embedded deep group convolutional network(EDGCN).And based on the two models respectively,build DGCN ensemble model and EDGCN ensemble model to improve the model performance score.In the ICU positive patient is alert guaranteed timely assistance.If the classification model predicts a false negative result,it will put the patient at risk.For comparison,we used a widely adopted index ‘Score’,which is defined as 100×(TP+TN)/(TP+TN+FP+5×FN),as performance evaluation.Compared with the traditional Accuracy indicator,Score specifically penalizes false negatives.Through the online test environment of the Physionet website,the best model in our research finally got 81.92,which is the second highest score among the reports so far.The contributions of this work are mainly reflected in:1 Our models are end-to-end,which means the original time series can be automatically mapped into a binary output,without manual feature extraction.2 The innovative adoption of group convolution makes full use of information from multi-channel signals.3 This work provides ideas for relating network parameter selection with physiological or physical natures of signals,especially the design of the convolution kernel size according to the time-scale characteristics of physiological signals,which provides a new idea for the design of network parameters for physiological signal analysis.4 Considering of the limited sample size,this work designs and compares a series of data expansion methods,which provides a way of thinking for the expansion of limited physiological time series in the future.