A Machine Learning Predictive Model for Sepsis

Sepsis Clinical Trial

Official title: A Machine Learning Model of Risk for Children With Sepsis

Status Recruiting

Clinical Trial Summary

Timely and accurately predicting the occurrence of sepsis and actively intervening in treatment may effectively improve the survival and cure rate of patients with sepsis.

Using machine learning and natural language processing, we want to develop models to 1) identify all children with sepsis admitted to hospital and 2) stratify them to distinguish those who are at high risk of death b) How will you undertake your work? From Shanghai hospitals anf MIMIC III, we will develop a very large dataset of patient admissions for all medical conditions including sepsis from the electronic health record. This data will include both structured data such as age, gender, medications, laboratory values, co-morbidities as well as unstructured data such as discharge summaries and physician notes. Using the dataset, we will train a model through natural language processing and machine learning to be able to identify people admitted with sepsis and identify those patients who will be at high risk of death. We will test the ability of these models to determine our predictive accuracies. We will then test these models at other institutions.

Clinical Trial Description

Introduction:Timely and accurately predicting the occurrence of sepsis and actively intervening in treatment may effectively improve the survival and cure rate of patients with sepsis. There have been a large number of research results on the prediction of sepsis produced by two methods: the sepsis detection and evaluation method based on clinical scoring mechanisms and the sepsis detection method based on machine learning model.

Objective: Reasonable and effective data pre-processing can significantly improve the timeliness and accuracy of early warning models of sepsis. Given the problems of high time dispersion, uneven distribution, and large differences of common sepsis prediction modeling indicators, the study proposed a method of hybrid interpolation based on time window-related sepsis indicators.

Methods:The study designed the traditional data interpolation method based on linear, MGP, average, nearest neighbour and the hybrid interpolation method (CTWH) based on correlation time window (CTW) proposed in the study for experimental comparison. Experiments were performed respectively in sample sets with no experimental data removal and sample sets with 90% missing values removal. By comparing with the performance of the existing sepsis indicator interpolation methods on the same baseline model, the effectiveness of the method was proven from the accuracy and timeliness of the prediction results. In the end, the results of the experimental method were analyzed and explained from a clinical perspective.

Significance:

In view of the characteristics of high dispersion, uneven distribution, and large differences between features of commonly used indicators in sepsis prediction models, this study proposed an efficient data interpolation strategy. After elimination of missing data of 90% and 0%, the interpolation method proposed in this study performed better than the existing methods like mean interpolation and linear interpolation, KNN, MGP on the static baseline and time series models. At the same time, this method also provided an idea to explore the length of the interpolation window, and supported the prospective study of missing value interpolation and data pre-processing. ;

Related Conditions & MeSH terms

• Sepsis
• Toxemia

• NCT number: NCT04771429
• Study type: Observational
• Source: Xinhua Hospital, Shanghai Jiao Tong University School of Medicine
• Contact: Qin Gao, MD
• Phone: 13761402225
• Email: 13761402225@163.com
• Status: Recruiting
• Start date: April 1, 2019
• Completion date: April 1, 2021