Medical informatics and machine learning for patient stratification and personalised prediction of outcome and clinical events in paediatric critical care

Background

Routine clinical practice generates a large amount of data that is under-used for research and quality improvement [1].  This is particularly true in paediatric intensive care units (PICU) across the world.  Every PICU patient routinely has multi-parameter bedside physiological monitoring data in at least minute-by-minute resolution collected throughout their PICU stay.  Yet once the patient is discharged, vital information from this physiological big data is discarded rather than being used to advance our understanding of how a patient’s physiological phenotype may affect outcome.

We urgently need to utilise data, leveraging recent advances in machine learning and explainable AI, to integrate and interrogate the wealth of data generated at bedside during routine patient care and develop precision medical approaches for critical care.  This information would enable clinical staff to continuously improve patient care and outcome and it could help generate new hypothesis for clinical research.

Our team has successfully defined clinical and physiological phenotypes associated with an improved global neurological outcome through the use of statistics and medical informatics [2,3]. We also have experience in the use of machine learning algorithms to preprocess multi-parameter physiological monitoring data.

Having established the principles of this project on physiological monitoring data, we now seek to make a step change and predict major clinical events, length of stay in critical care, and outcome in critically brain injured paediatric patients, and to explain the main factors behind them.

A key feature of our approach is the focus on integrating information from diverse physiological monitoring recordings.  This is very timely given recent research on the promise that integrating multi-parameter data with machine learning would bring to the field [4].  Yet, the application of machine learning in this domain remains in its infancy as it has focused on single parameter vitals.  This narrow focus overlooks the rich tapestry of information available in multi-parameter data to better understand recovery and outcome in paediatric critical care at an individual patient level [4].  Our proposition will lead to personalised therapeutic strategies and has the potential to be extended to include all patients in critical care regardless of disease and age.

Aims

This interdisciplinary project aims to develop data informatics and explainable machine learning algorithms for data driven analysis of routinely collected paediatric critical care data. We hypothesise that clinical and physiological phenotypes in paediatric patients affect outcome and that the developed algorithms will be able to obtain data driven representations of such cases useful to predict clinical events and length of stay in the unit.

Training Outcomes

We provide a unique opportunity for the student to apply their developing quantitative and interdisciplinary skills to a variety of data types and sources. Specifically:

Digital Excellence

•         Understanding of the data collection environment in intensive care units.

•         Experience with diverse, big data.

Quantitative

•         Supervised and unsupervised machine learning algorithms.

•         Medical informatics.

•         Explainable AI.

•         Analysis of physiological time-series routinely collected at clinical bedside.

Interdisciplinary

•         Domain Knowledge in paediatric critical care pathophysiology (clinical / physiological ‘phenotypes’).

•         Development of personalised precision medicine for paediatric critical care.

References

[1] Celi, et al. “Big data” in the intensive care unit. Closing the data loop. Am J Respir Crit Care Med 2013. 187(11):1157-1160.

[2] Güiza, et al. Continuous optimal CPP based on minute-by- minute monitoring data: a study on a pediatric population. Acta Neurochir 2016. 122:187-191.

[3] Guiza et al. Visualizing the pressure and time burden of intracranial hypertension in adult and paediatric traumatic brain injury. Intensive Care Medicine 2015. 41(6):1067-1076.

[4] Foreman et al. Challenges and Opportunities in Multimodal Monitoring and Data Analytics in Traumatic Brain Injury. Curr Neurol Neurosci Rep, 2021 ;21(3):6.