20 Predicting paediatric sepsis through vital dynamics and machine learning. (23rd February 2023)

• Record Type: Journal Article
• Authors: Papanastassiou, Helen; Bowyer, Stuart A; Booth, John; Briggs, Lydia; Bryant, William A; Key, Daniel; Shah, Mohsin; Spiridou, Anastassia; Sebire, Neil J
• Is Part Of: Archives of disease in childhood. Volume 108(2023)Supplement 1
• Page Start: A7
• Page End: A8
• Publication Date: 2023-02-23
• Subjects: Children -- Diseases -- Periodicals; Infants -- Diseases -- Periodicals; 618.920005
• DOI: 10.1136/archdischild-2023-gosh.20
• Languages: English
• ISSNs: 0003-9888

Abstract

Abstract : Sepsis is a complex clinical syndrome that, without early clinical intervention, can lead to septic shock, organ failure and death. It is one of the leading causes of paediatric death worldwide. With no clear biomarker to indicate its onset, researchers have looked to vital sign dynamics to provide novel information for the period preceding a septic episode. With the increasing availability of electronic patient records (EPR) and mechanisms to analyse big sets of data, such as machine learning algorithms, there is a potential to develop clinical support tools that can warn clinicians before an episode. We investigated the efficacy of a sepsis classification approach developed by Bloch et al. based on the variability in four routinely recorded vital signs. The features were extracted from vital sign recordings of patients who were admitted to the paediatric ICU of Great Ormond Street Hospital and had a sepsis diagnosis during their stay. We tested four machine learning algorithms in a classification task, attempting to distinguish between non-sepsis patient days and sepsis patient days. After encountering a class imbalance in our dataset (5974 negative cases and 153 positive cases), four variations of the Synthetic Minority Oversampling Technique (SMOTE) were employed on our training dataset to improve our model's learning ability. The four machine learning algorithms, after oversampling, achieved precision scores up to 12.5% (SVM and SMOTE & ENN), and recall scores up to 64.5% (Logistic Regression and SMOTE). Random Forest with no resampling achieved an AUC-ROC score of 80.2%. Future research should test different hypotheses to establish the role of EPR data in predicting paediatric sepsis and explore how the false positive rate can be reduced to make the method more suitable as a clinical support tool.