Research On Sepsis Auxiliary Diagnosis With Deep Learning

Sepsis is a critical illness caused by infection and is one of the leading causes of patient death in the ICU.Sepsis progresses rapidly and the effective resuscitation time is short,which lead to the the lower the survival rate without timely diagnosis.The key to patient care is timely diagnosis and interventional treatment.In this paper,researches on predictive diagnosis of early sepsis under different scenarios are carried out based on deep learning and machine learning techniques.Different research tasks are proposed according to the clinical demands,which mainly include: sepsis prediction for patients based on the initial examination data of admission,real-time sepsis prediction in stages based on partial test data,and sepsis prediction based on the temporal monitoring data during patients’ hospitalization.We propose the corresponding deep learning-based network models for each of these problems and obtain satisfactory performance.In this paper,a cross feature-based deep encoding network framework(CF-DEN)is proposed for the sepsis prediction task using medical records of initial examination after admission.The feature cross module automatically constructs higher-order cross features by using paths and nodes inside the tree model and feeds them into a deep encoding network(DEN).The DEN network layers filter the features involved in the current layer computation and learn nonlinear features through an attention mechanism.Each network layer outputs the current embedding features for layer-by-layer accumulation to obtain the embedding features in the final layer.The framework integrates the tree-based approach and the neural network approach to efficiently handle small clinical datasets and obtain accurate prediction performance.In this paper,the performance of the framework is evaluated on a dataset from Shanghai Children’s Medical Center.The AUC and f1-score of the model are significantly improved compared to other algorithms,which can provide more accurate diagnosis for patients.We propose a staged sepsis prediction model for real-time diagnosis with graph neural networks and ensemble methods.The six admission examinations are divided into three stages according to the length of the test items.A predictive model based on partial information is proposed to shorten the time for sepsis patients to receive their first antibiotic treatment intervention.The model is trainsed in a self-improvement manner,and the prediction accuracy gradually improves with data refinement.In this paper,the performance of the model is evaluated on a dataset collected at the Shanghai Children’s Medical Center.The f1-scores of the real-time predictive model are 77.35%,85.71%,and 86.48% when using the first stage test data,the second stage,and the third stage,respectively.The model obtains relatively high prediction accuracy with a shorter diagnosis time span in the second stage,which may help to reduce patient waiting time.We propose a sepsis sequential predictive model for positive and negative prediction of sepsis for each time interval since admission to the intensive care unit based on the sequence of electronic health records of patients during their hospitalization.The model is divided into two stages: multidimensional fusion feature construction and fast decision making.The model constructs multidimensional fusion features by extracting mapping vectors from time-series data through a path signature algorithm and a sequence feature transformer.The sequence feature transformer projects current input data and historical data memory into a high-dimensional embedding vector.The path signature algorithm treats the physiological sequence data of patient clinics as paths and computes signatures by path cumulative integration to provide geometric changes of data along different directions.The fast decision forest model is proposed as a classifier in the fast decision phase to give the final prediction.The sepsis sequence diagnosis model achieves high prediction accuracy on two hospital datasets,which helps hospitals to achieve real-time monitoring and early warning of patients’ physiological status.