基于机器学习算法预测脓毒症性凝血病患者ICU住院时间OA

Objective To study the construction of a predictive model for prolonged ICU stay in patients with Sepsis-in-duced coagulopathy(SIC)using six machine learning methods,analyze its related characteristic risk factors and clinical signifi-cance,so as to identify SIC patients with prolonged ICU stay at an early stage,provide patients with more accurate clinical treat-ment plans,and improve the utilization rate of ICU medical resources.Methods A total of 3 728 patients who met the diagnos-tic criteria for SIC were screened from the Medical Information Database for Intensive Care(MIMIC-Ⅳ).According to the third and quartile values of the ICU stay of all patients in the queue,the SIC patients were divided into the prolonged ICU stay group(≥ 5 days)and the non-prolonged ICU stay group(＜5 days).Collect their general data,clinical data and laboratory test re-sults within 24 hours after admission to the ICU,and analyze the independent risk factors for prolonged ICU stay in patients with SIC.The predictor variables were jointly screened through the minimum absolute contraction selection operator LASSO-Logistic regression.The screened predictive variables were respectively used to construct six machine learning models,namely Random Forest(RF),Extreme Gradient Boost(XGBoost),Logistic Regression,Decision Tree(DT),K-nearest neighbor(KNNC),and Multi-layer perceptron(MLP),to predict the prolonged ICU stay of patients.The model performance was evaluated by using the receiver operating characteristic(ROC)curve,calibration curve and clinical decision curve(DCA).And the interpretability analysis of the simplified optimal model is conducted using Shapley Addition Interpretation(SHAP)and constructed Nomogram.Results A total of 3 728 patient samples were included in this study,among which 832 patients had an ICU stay of ≥5 days and 2 896 patients had an ICU stay of＜5 days.Twelve predictive variables,including age,SOFA score,heart rate,white blood cell count,red blood cell distribution width,international normalized ratio of INR,percentage of lymphocytes,percentage of monocytes,complications of acute kidney injury,complications of heart failure,vasoactive drug treatment,and mechanical venti-lation treatment,were screened out based on LASSO-Logistic regression.Among the six machine learning models,namely Ran-dom Forest(RF),Extreme Gradient Boosting(XGBoost),Logistic Regression,Decision Tree(DT),K-nearest neighbor(KNNC),and Multi-layer perceptron(MLP),the areas under the receiver operating characteristic curves are as follows in se-quence:0.743,0.705,0.704,0.678,0.636,0.688.Among them,the predictive efficiency of the RF model is the best.Meanwhile,the accuracy rate of the RF model is 81.23％,the sensitivity is 60.00％,and the specificity is 96.87％.Overall,compared with other models,the RF model can better predict the prolonged ICU stay of patients,and has better accuracy.The order of importance of the top three characteristics in SHAP is:SOFA score,complications of acute kidney injury and mechanical ventilation treatment.The Nomogram was constructed to make this study more clinically applicable.Conclusion The prediction model for SIC patients constructed based on the Random Forest(RF)algorithm can effectively predict the possibility of prolonged ICU stay for patients.The explanatory analysis provided by SHAP and Nomogram can support clinical decision-making,which is helpful for clinicians to intervene in patients with sepsis coagulation at an early stage,shorten the length of ICU stay,and improve the prognosis of patients.