Machine Learning - Early Warning System (ML-EWS)

Patients admitted to hospital can get better, but also sometimes become unexpectedly worse, facing the need for emergency resuscitation, surgery, intensive care unit admission, or even death. These are collectively categorised as a Serious Adverse Event (SAE). Stopping this happening requires identifying those at risk, and a lot of effort has been made to do this. However, past efforts have failed to produce a working system, often because they have used limited information obtained only on admission to hospital, or because expensive new equipment has been needed in order to make them work.

I will address this problem by developing a software system, a simple and widely- applicable analysis tool, for continuous tracking of SAE risk in all patients admitted to hospital. This system will have several unique advantages. Firstly, it will apply to all hospitalised patients and will track the risk of an SAE as it changes over time. Secondly, it will use routinely- collected information from blood tests, married to existing patient data relating to their other illnesses, personal details and medical treatments and tests. My preliminary data suggests that the result will be an accurate system which predicts risk, and which is cheap and easy to implement in all hospitals. This will help clinicians make the right decisions on further investigations, referral to specialist teams and specific treatment earlier than what currently occurs. This allocation of the right resources at the right time will hopefully help prevent unnecessary suffering and death in hospital.

Serious adverse events (SAEs, such as cardiac arrest, unexpected intensive care unit admission, the need for emergency surgery, or death) are commonplace in hospitalised patients. A key factor in reducing the burden of unexpected in-hospital morbidity and mortality is the identification of those at risk of deterioration. Previous attempts to do this have focused only on limited variables (often at one timepoint), have had poor sensitivity/specificity, and/or have necessitated the introduction of expensive new technical systems and management processes. As a result, no effective system has yet been implemented. I therefore propose to develop a simple and widely- applicable multivariate decision tool for continuous tracking of SAE risk in all patients admitted to hospital. This will use high- dimension multivariate analysis of routinely collected blood tests will be married to existing patient data relating to co-morbidities, interventions and demographics. Little additional resource would be required to collate the data, interpret and act upon it, as all of these data are already stored electronically. If successful, this research will permit automatic and continuous individual risk stratification to guide clinical decision making and resource allocation, ultimately helping prevent serious adverse events.

The proposed research will demonstrate the scientific utility of an automated analytical system that combines existing data from multiple hospital systems. Specifically, it will use routinely- collected information from blood tests, married to existing patient data relating to their other illnesses, personal details and medical treatments to track patient illness trajectory. It will identify, at an early stage, those at risk of deterioration, thus enabling early intervention to prevent this.

In-Hospital Patients
A reduction in in-hospital morbidity, mortality and an increase in the quality of care.
- Mapping an individual patient's illness trajectory allows for real time forecasting of serious adverse events with a much higher sensitivity and specificity than current techniques.
- Early detection of deterioration will allow for early medical intervention by specialist medical teams.
- Early intervention via medical emergency teams has been shown to reduce both cardiac arrests and death.
Early detection and intervention is a recommendation of both NICE and NCEPOD.

Clinicians
- Informing on the health status and trajectory of their individual patients, automatically and immediately, without the need for additional investigations.
- Providing a real time comparison of their individual patient's risk of morbidity and mortality to all other patients in their hospital currently as well as historically.
- Demonstrating the utility of a large sale data driven and low cost approach to improving healthcare.

National Health Service
- Automated and continuous monitoring of all in-hospital patients: An IT system tailored for in-hospital use that is easy to deploy, maintain and has a minimal cost.
- Better resource utilisation: understanding the trajectory of patient illness with an accurate early warning system will allow for prediction of level of care required.
- Automated interpretation of physiology could aid in the triage of all attendances to hospital.

Medical Research Community
- The generation of a unique but vitally important dataset unparalleled in detail. This has heritability beyond the scope of the proposed project and will contribute to the future development of improved systems and methods.
- Patterns of physiological changes identified as precursors to patient deterioration could open new avenues of research into understanding underlying pathology and the effectiveness of interventions.

Mathematical Research Community
- The proposed analyses build on state-of-the art generative models of time-series & constitute significant research agendas on their own.
- The scale of data collected requires the development of robust and fast online inference procedures that are able to provide both intuitive summary statistics of the state of the patient and also predictions as to the likely outcome of the patient. Developing such techniques will push the application of approximate inference techniques in Bayesian time-series models.

Intellectual property
The core intellectual property of the project would reside in the algorithms developed through a combination of clinicians' input and machine learning.
- Software (Copyright): new software implementation of prediction models.
- Software (Patent): novel diagnostic methods for analysis of derangement of blood results & patient data.
- Database Rights: blood result/ patient data/ clinical outcome databases.

Health Technology Company
A new medical technology that implements the IT solution, either through licensing to existing NHS IT providers (e.g. Cerner) or through the development of an independent platform.

Economy & Society
Better utilisation of resources coupled with decreased morbidity & mortality will provide better value for money for the taxpayer while at the same time reducing the illness burden on the economy.

Dr Vishal Nangalia
Development of advanced interdisciplinary skills in machine learning & generic research skills.