SAMueL – Stroke Audit Machine Learning

What is the SAMueL-1 project?

The SAMueL-1 project team developed modelling tools to understand the variation in the use of clot-busting drugs in hospitals, and support services to optimise their use.

What is the background?

Stroke is a leading cause of death and disability with over 85,000 people hospitalised in the UK each year. One way of treating stroke and preventing disability is to give a patient medication to break down blood clots. This is called thrombolysis.

However, thrombolysis is not suitable for all patients and can be risky. For thrombolysis to be useful it needs to be given as soon after the stroke as possible. Use of thrombolysis varies hugely, even for patients with similar treatment pathways and with similar characteristics; some hospitals rarely use it while some use it in a quarter of stroke patients. The speed of giving thrombolysis also varies, with some hospitals taking an average of 90 minutes and others taking less than 40 minutes to administer the drug.

How are researchers trying to help?

Using modern computer science techniques of clinical pathway modelling and machine learning, we aimed to find out why there was so much variability in the use of thrombolysis. This helped hospitals understand what they can do to optimise its use.

Based on the decisions made by highly qualified stroke experts we built a tool for assisting doctors to review their use of thrombolysis. This was particularly useful for smaller hospitals, without access to as experienced staff, to compare their decision-making with hospitals with easier access to specialist and experienced stroke experts.

What has happened so far?

The research team used a state-of-the-art computer modelling technique – pathway modelling – to better understand what causes variation in care across the UK. This approach replicates, in a computer model, the flow of patients through the first few hours of stroke care, mimicking the same processes and timings that the stroke unit currently provides. This allowed us to look at the effect of changing key aspects of patient flows in a controlled, modelled environment, without affecting real patients. A second technique called ‘machine learning’ enabled us to teach a computer the likely decision made in any hospital given any particular patient.

Both approaches allow us to ask ‘what if?’ questions, such as ‘what if a hospital improved diagnosis of patients by asking more questions, but by doing so extended patients’ waiting times for scans?’. With machine learning we can ask ‘what if the decisions at all hospitals were similar to hospitals that are considered centres of excellence for stroke care?’. By asking these types of questions we can identify changes at each hospital of most benefit to patients. Both techniques have been piloted across seven hospitals.

Through this study we found that the factors that would most improve thrombolysis use were aiding clinical decision making, more precise knowledge of when a stroke occurred and speeding up the stroke pathway at hospitals.

We would now like to test and refine these methods across all stroke units in England. A researcher will interview doctors to understand their attitudes to thrombolysis, and how the results from the modelling work can best be presented to them in a way that will influence more consistent stroke care across the UK.

We conducted this work with the National Stroke Audit, hosted by the Royal College of Physicians. Our aim was to build these new advanced analytic tools into the quarterly stroke audit, helping hospitals understand whether their use and speed of thrombolysis is different from that expected for their patient population, and what changes would most improve performance (if needed).

The work has been cited in the last national clinical guidelines for stroke, on understanding achievable thrombolysis rates across hospitals in England.

What happens next?

Apply for funding to:

– Build ‘Explainable Machine Learning’ models to understand what factors affect the decision to give thrombolysis or not at different hospitals.

– Expand the machine learning to predict the probability of a good outcome, and probability of adverse effects of thrombolysis, and to understand what patient features most influence outcomes after stroke, with and without thrombolysis.

– Conduct further qualitative research with clinicians from different hospitals with a focus on most useful outputs from this work for clinicians

– Explore the cost-effectiveness of making organisational changes to the care pathway, demonstrating benefits in health economic terms such as Quality Adjusted Life Years

– Include organisational features to address whether hospitals have insufficient infrastructure for an effective stroke pathway which might strengthen the argument for more investment in the stroke pathway