Minh Nguyen, MS
PhD Student
National University of Singapore
Singapore, Singapore
Disclosure information not submitted.
Kim Leng Poh, PhD
Associate Professor
National University of Singapore, Singapore
Disclosure information not submitted.
Shu-Ling Chong, MBBS, MRCPCH, MCI, MPH
Clinical Associate Professor
KK Women's and Children's Hospital, Singapore
Disclosure information not submitted.
Lee Jan Hau, MBBS, MRCPCH, MCI (he/him/his)
KK Women's and Children's Hospital, Singapore
Singapore, Slovenia
Disclosure information not submitted.
Title: Effective Diagnosis of Sepsis in Critically Ill Children Using Probabilistic Graphical Model
Introduction: Early sepsis diagnosis in children remains a challenge for clinicians. To improve the diagnosis, Probabilistic Graphical Model (PGM), a rich framework that uses graphs to model uncertain associations between a large number of variables in complex domains, can be utilized. However, its application in pediatric sepsis remains limited. This study aims to explore the utility of PGM in pediatric sepsis by evaluating the diagnostic capability of variables and biomarkers commonly used in the pediatric intensive care unit (PICU).
Methods: The publicly available Paediatric Intensive Care (PICD) dataset was utilized for this study. We included children with vital signs and clinical data documented within 24 hours of admission and a discharge diagnosis of sepsis (defined by ICD-10). A PGM method, Tree Augmented Naive Bayes (TAN), was used to build models using combinations of 4 categories: vital signs, clinical symptoms, laboratory tests, and microbiological tests. Variables were reviewed and selected by clinicians. Data were split into 70-30 for training and testing, respectively with 10 k-fold cross-validations. Diagnostic performance was measured by sensitivity (SEN), specificity (SPE), accuracy (ACC), and Area Under the Curve (AUC).
Results: A total of 12,826 patients (age: 2.4 ± 3.6 yrs.; 41.3% female) were included from 2010 – 2019. Of these, 368 patients were diagnosed with sepsis (2.9%) with an overall mortality rate of 12.5%. All models had high SPE (0.78-0.99) and AUC (0.67-0.84). Models that utilized more than two categories performed better than those that used a single category. The combination of all categories provided the strongest overall performance (SEN: 0.72, SPE: 0.8, AUC: 0.84, ACC: 0.8). Laboratory tests were efficient in detecting sepsis (SEN: 0.67-0.72). Microbiology tests had low SEN (0.3) and high SPE (0.99) due to the high incidence of negative results (30%).
Conclusions: We demonstrated that PGM is an efficient diagnostic tool for pediatric sepsis. More studies should be conducted to assess its utility to aid clinicians in the diagnosis of sepsis within the PICU