Predicting patient deterioration with physiological data using AI: systematic review protocol

Introduction

Patient observation has been an integral part of clinical practice since antiquity. Florence Nightingale strongly advocated structured patient observations in the second half of the 19th century,1 and healthcare professionals’ systematic recording of vital signs in the UK was formalised from the 1950s.2 These repeated sets of physiological measurements are used to assess the severity of illness, and in 2012 the Royal College of Physicians introduced the National Early Warning Score (NEWS) to create a standardised approach to recognise and respond to patient deterioration.3 4 This was updated in 2017 as NEWS2 with changes including a dedicated section for patients with hypercapnic respiratory failure.5 NEWS and NEWS2 have been shown to reduce mortality,6 7 but have limited positive and negative predictive accuracy.3 8 9 A recent study showed that at the clinically significant 5-point threshold, NEWS2 has a sensitivity of 72%, a specificity of 86% and a number needed to evaluate of 8.2 for the composite of cardiac arrest, unplanned critical care admission or death within 24 hours of the observation.10

Since then, digital tools to record observations (e-Observations or eObs) have been adopted widely. NEWS2 alone is now used by 100% of ambulance trusts and 76% of acute trusts across England, UK.11 The wide adoption of digital tools provides an opportunity to use artificial intelligence (AI) tools to interrogate these routinely collected data and improve the predictive accuracy of NEWS2. Several studies have demonstrated the capability of machine learning (ML) models in predicting clinical deterioration, particularly in emergency departments.12–14 By using ML and other AI tools to monitor and interpret physiological data, it is possible to include a wider set of variables than humans would have in a healthcare setting. Incorporating additional variables derived from electronic health records was beneficial in improving the predictive accuracy of the scoring system than using just vital signs.15 16 Our recent scoping review evaluated potential additional variables that may be used to improve the predictive accuracy of NEWS2.10 This systematic review investigates the AI approach that would be most appropriate for interrogating these data. We also aim to establish whether the same AI approach would be suitable for additional variables such as age, comorbidity and prescribed medications to provide a more personalised NEWS. While there are a few existing systematic reviews about ML models for the prediction of clinical deterioration, they focused on the barriers to the implementation of existing models in hospital settings or the evaluation of model performance.17 18 What still needs to be added to the literature is how to choose an ML-based model that is accurate and suitable for predicting disease deterioration with the current patient data structure in hospitals. This systematic review aims to contribute to the existing body of evidence and provide a reference to promote the implementation of ML-based models in hospital settings.

Back to top

Methods

The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA; see online supplemental appendix A for the PRISMA checklist)19 and PICOS (Population, Intervention, Comparator, Outcome and Studies) frameworks20 were used to build the search strategy (table 1) and provide a framework for the review.

Back to top