Prediction Of In-hospital Mortality Risk In Intensive Care Unit Based On Machine Learning

Objectives:Intensive Care Unit(ICU)is established for the treatment of severe patients with advanced medical equipment and professional medical care personnel.It is committed to providing the optimal treatment for severe patients.Due to the serious illness or trauma,the patients in ICU have a high in-hospital mortality risk.Worldwide,the mortality in ICU was11.3% in the United States as high as 41-44% in some parts of Europe,and roughly 16.3% in China.The mortality of severe patients varies greatly with the evolution of the underlying disease.The mortality of the patients admitted to the ICU after elective surgery could be as low as 5%,while that may be 25% for patients with respiratory diseases.With the outbreak of COVID-19 in 2019,more and more critically ill patients were transferred to ICU,bringing great pressure to the global health system and resulting in the increasing mortality in ICU.Accurate prediction of mortality risk is helpful for clinicians to identify patients with high mortality risk as soon as possible,so that clinical intervention could be given as soon as possible and the nursing grade could be adjusted properly.We can obtain the result of improved survival and reduction of mortality for critically ill patients.At present,the widely used assessment tools are difficult to meet the increasing requirements for medical data analysis.Therefore,it is urgent to develop new tools to deal with the problem.Our study established a model for the rapid,accurate,and static prediction of in-hospital mortality risk in ICU based on support vector machine(SVM)with non-time series data from the Medical Information Mark for Intensive Care(MIMIC-III).Moreover,we developed an attention-based TCN model to dynamically predict the mortality risk of ICU patients with time series data.Our study aims to explore new methods for the prediction of mortality risk in ICU.Methods:1.Data used for the prediction of mortality risk were extracted from the ICU critical illness database,MIMIC-III.We got access to the database after applying on its web page and passing its assessment.We screened the MIMIC-III database according to inclusion and platoon criteria.And 18,094 patients were included in the total data set.The patient’s physiological variables used in our study were determined according to the principle(referring to those that can effectively reflect the disease state and treatment effect).Then data24 h and 48 h after ICU admission were extracted and washed.We calculated the statistical characteristics(non-time series data)of every physiological variable of each patient over the same time period.We split the dataset into training set(15331 patients)and testing set(2763patients).2.After selecting features on the training set using Lasso regression,we developed an SVM-based model to statically predict the mortality risk 24 h and 48 h after ICU admission using the Averaged Neural Network(avNNET)and evaluated the effect.According to the confusion matrix,we calculated and compared sensitivity,specificity,accuracy and the area under the ROC curve(AUC).3.We developed an attention-based TCN model to predict the mortality risk of ICU patients with time series data from the total data set and compare the performance of the attention-based TCN model and non-time series models based on Logistic Regression(LR),Support Vector Machine(SVM),Random Forest(RF),Decision Tree(DT)by calculating and comparing recall,specificity,F1-score,Brier score,AUCROC and the area under the Precision-Recall curve(AUC-PR).Result:1.On the training set,the AUC of the models for statically predicting the mortality risk24 h and 48 h after ICU admission was 0.840(95% CI: 0.835-0.844)and 0.829(95% CI:0.823-0.832),with sensitivity of 81.24% and 77.62%,specificity of 74.15% and 74.51% and accuracy of 77.66(95% CI: 0.772,0.781)and 76.05(95% CI: 0.756,0.765).On the testing set,the AUC of the models for statically predicting the mortality risk 24 h and 48 h after ICU admission was 0.805(95% CI: 0.794-0.817)and 0.811(95% CI: 0.800-0.824)with sensitivity of 75.13% and 73.72%,specificity of 71.30% and 74.29%,and accuracy of 71.82%(95% CI:0.709-0.727)and 74.21%(95% CI: 0.733-0.751).2.As for the model for dynamically predicting the mortality risk of ICU patients,the attention-based TCN model has slightly lower AUCROC and AUC-PR than the non-time series ML methods.But the former had the highest sensitivity(67.1%)and F1 score(0.46).It’s Brier score(0.142)was also higher than other traditional ML models.Compared with other time series methods,the sensitivity of the attention-based TCN is much higher than that of the LSTM model(46.1%)based on the same database,and the AUC-PR between the two models differs very little.To sum up,The attention-based TCN had high accuracy and low missed diagnosis rate and performed the best among the models.Conclusion:1.The study to establish an SVM-based model of statically predicting the in-hospital mortality risk in intensive care unit with non-time series data is practicable.The model has good predictive effect,satisfactory stability and a high accuracy.2.The attention-based TCN model achieved better performance in the prediction of mortality risk with time series data than non-time series models,which suggests that it is feasible to dynamically predict in-hospital mortality risk in ICU for helping decision-making for critical patients with continuous data flow.