Related Experiment Video
Updated: Oct 26, 2025
A Neonatal Imaging Model of Gram-Negative Bacterial Sepsis
Published on: August 12, 2020
Towards real-time diagnosis for pediatric sepsis using graph neural network and ensemble methods
1University of Chinese Academy of Sciences, Beijing, China. nalanyu2000@163.com.
Insights
This study introduces a real-time pediatric sepsis prediction model, significantly reducing diagnosis time. The model achieves high accuracy even with initial data, enabling faster antibiotic treatment for critically ill children.
Area of Science:
- Pediatric critical care medicine
- Artificial intelligence in healthcare
- Sepsis diagnostics
Background:
- Pediatric sepsis presents a rapid threat in ICUs, necessitating swift resuscitation.
- Current sepsis prediction lacks research for short time intervals, delaying critical treatment.
Purpose of the Study:
- To develop a predictive model for real-time sepsis diagnosis in pediatric intensive care units (ICUs).
- To reduce the time to the first antibiotic treatment for pediatric sepsis patients.
Main Methods:
- Utilized data from Shanghai Children's Medical Center, including medical history, physical exams, and six lab test groups.
- Employed a graph neural network for real-time feature extraction and a deep forest model for comprehensive prediction across three data stages.
- Integrated discriminative features from previous stages to optimize global judgment.
Main Results:
- Achieved high Area Under the Curve (AUC) scores: 93.63% (stage 1), 96.73% (stage 2), and 97.58% (full data).
- Obtained strong F1-scores: 77.35% (stage 1), 85.71% (stage 2), and 86.48% (full data).
- Demonstrated accurate predictions at each stage, with stage 2 accuracy approaching full data results.
Conclusions:
- The real-time sepsis prediction model offers superior accuracy compared to control methods.
- Early data input (first two stages) provides near-complete data accuracy, compressing diagnosis time to approximately one hour.
- The model facilitates reduced waiting times and aids in proactive medical resource allocation for sepsis management.
Objective:
The rapid onset of pediatric sepsis and the short optimal time for resuscitation pose a severe threat to children's health in the ICU. Timely diagnosis and intervention are essential to curing sepsis, but there is a lack of research on the prediction of sepsis at shorter time intervals. This study proposes a predictive model towards real-time diagnosis of sepsis to help reduce the time to first antibiotic treatment.
Patients And Methods:
The dataset used in this paper was obtained from the pediatric intensive care unit of Shanghai Children's Medical Center and consisted of the initial examination records of patients admitted to the hospital. The data included six groups of laboratory tests: medical history, physical examination, blood gas analysis, routine blood tests, serological tests, and coagulation tests. We divided the admission examination into three stages and proposed a sepsis prediction model towards real-time diagnosis based on local information to shorten waiting time for treatment. The model extracts homogeneous features from patient groups in real-time using a graph neural network and uses the deep forest to learn from homogeneous features and laboratory data to give a comprehensive prediction at the current stage. Discriminative features of each stage are used as augmented information for the next phase, finally achieving self-optimization of global judgment, assisting in pre-allocation of medical resources and providing timely medical assistance to sepsis patients.
Results:
Based on the first stage, second stage, and full test, the AUCs of our model were 93.63%, 96.73%, and 97.58%, respectively, and the F1-scores were 77.35%, 85.71%, and 86.48%, respectively. The models gave relatively accurate predictions at each stage.
Conclusions:
The prediction model toward a real-time diagnosis of sepsis shows more accurate predictions at each stage compared to other control methods. When the first two stages of data are obtained as input, the model accuracy is close to using complete test data, which can help compress the time to diagnosis to about an hour after the test and significantly reduce waiting time.
