Designing a Paediatric Sepsis Prediction Tool using Machine Learning

BACKGROUND: Sepsis in paediatrics is a global threat and a major contributor to childhood mortality and morbidity. Recognition of sepsis was acknowledged as a national, and also a top research priority.

There are limited studies reporting the utilization of paediatric sepsis prediction tools, none of which have been subjected to robust research trials to allow generalisability. To date there are no validated paediatric sepsis prediction tools in the UK.

Models developed utilizing electronic health records (EHR) using machine learning (ML) algorithms, demonstrated high precision to detect neonatal sepsis and paediatric severe sepsis, and predict clinical outcomes and disposition in children in the emergency department (ED).

AIM: The overarching aim of the study is to design a robust, accurate prediction tool for identifying the outcomes of sepsis in children, using machine learning methods. PRIMARY OBJECTIVE: Derive clinical characteristics predictive of sepsis outcomes using ML methods.

SECONDARY OBJECTIVE: Create a ML algorithm to predict sepsis outcomes METHODS: The study designed as a retrospective, observational database analysis, will curate a 2-year dataset of routinely collected EHRs of children less than 16-years, attending the paediatric ED at the St Thomas' Hospital.

PRELIMINARY WORK: A sample dataset of 1199 patients with infection (fever) was curated for six variables (age, heart rate, respiratory rate, oxygen saturation, systolic blood pressure, AVPU scale). We trained three classifiers - a logistic regression, a random forest and XGBoost model to predict admission.

We split the data into a 75:25 ratio for training and testing respectively. We trained each model with the training data and evaluated the AUROC score on each data split.

The performance of these models (training vs test AUROC) were logistic regression 0.74 (0.73), random forest 0.9 (0.77) and XGBoost Classifier 1.0 (0.74). Each model performed better than random guess (AUROC > 0.5) and had a much higher accuracy. We demonstrated predictive power in the six variables.

Incorporating more clinical variables may enhance the accuracy.

The reasonable performance of this model demonstrated this is a viable approach and the overfitting justifies the need for further data collection.

STRATEGY: 1) Develop the infrastructure and platforms for the extraction, storage and processing of a wider range of EHR variables, including laboratory and microbiology results, 2) Create ML prediction model using classical statistical and machine learning methods, combining novel predictors of sepsis with criteria from known scoring systems of paediatric sepsis 3) Train, test and improve machine learning model.

SIGNIFICANCE: Our study will pioneer the application of machine learning for the first time in paediatric emergency care in the UK.

Further, it will initiate important research using big data and artificial intelligence in paediatrics and other specialties. The project fosters collaboration of clinical experts, data analysts and data scientists.

The ML algorithm developed during the study will undergo a Phase 2 multi-centre prospective evaluation and refinement to demonstrate its clinical effectiveness, and its broader therapeutic and economic impact.

The final prediction tool will have the potential to save children s lives and improve outcomes from sepsis in children though nationwide adoption.