Early Prediction Model Of Sepsis Based On Machine Learning Algorithm

Smart healthcare is based on big data to provide systematic,precise and precise medical services with artificial intelligence technology.The current main development directions are gene sequencing,computer aided diagnosis,medical imaging and drug research and development.With the rapid development of big data and AI technology,the medical field is steadily moving towards the stage of smart medical care.In response to the development trend of smart medicine,computer aided diagnosis of AI in early sepsis will be primarily focused in this paper.Sepsis is organ failure caused by infection,a critical illness with extremely high mortality.Early detection of sepsis and antibiotic treatment are critical to control the disease.A machine learning model with high predictive performance and clinical interpretability is aimed to be trained to predict early sepsis 6 hours in advance.Physiological data in this paper are from three independent hospitals,used into two processing programs called simple model and a full model,in view of the serious lack of data and the imbalance of categories.In the simple model,this article creatively divides the physiological state of patients with sepsis into a safe period,an early warning period,and a diseased period,and unifies the physiological state of non-septic patients as a safe period,extracting from the different states respectively mean vectors and the corresponding label of whether you suffer from sepsis in the next 6 hours;in the full model,sepsis patient data through under-sampling are extracted,and statistical strength features,window features,and medical features are constructed based on the original variables Index characteristics.Then,the data processed by the above two methods are used to fill in the missing values by the miceforest multiple interpolation method,and the XGBoost and Light GBM algorithms are used for predictive analysis,and the performance of the model is comprehensively evaluated,combined with SHAP Values to explain the classification logic of the model.The results show that: the performance of the full model is better than the simple model in Accuracy,Recall rate,Kappa coefficient,etc;the performance of the two algorithms is excellent in prediction,among which Light GBM has faster running speed and stronger generalization ability in multidimensional data,with AUC reaching 0.979 in the full model;PTT(thrombin-activating enzyme time),WBC(white blood cell count),Platelets are the key risk factors for predicting early sepsis.The innovations of this paper are: 1)In the simple model,the solution of extracting the mean basis vector is adopted to divide the time window of the early warning period every 2hours or 3 hours to capture the change characteristics of each variable in the early warning period,and the obtained model performance is compared with Compare the original model;2)In the full model,the forward filling method and the miceforest interpolation method are combined to fill the data,and the model results show that the filling effect is good.