Back to blog

How can we use AI to predict recovery in stroke?

Dr. Smriti Agarwal, Consultant Neurologist, Department of Stroke Medicine, Addenbrooke’s Hospital; Visiting fellow, Victor Phillip Dahdaleh Heart & Lung Research Institute (HLRI)/Cambridge Mathematics of Information in Healthcare (CMIH) Hub, University of Cambridge

28 November 2024

Stroke is a leading cause of adult death and disability in the UK and worldwide. Over 80% of strokes are due to blood clots in the arteries supplying the brain. Treatments to open up blocked arteries, by an injection of a clot-busting drug (thrombolysis) or a procedure to remove clots (thrombectomy) soon after a stroke, have transformed what we can offer to patients. However, despite effective treatments, 1 in 4 patients in clinical trials recover poorly from their stroke. Predicting an outcome for an individual patient relies on a number of factors, including severity of stroke and what their brain looks like on a scan.

Working with Professor James Rudd and Professor Carola Schönlieb at CMIH, I co-lead a team of researchers developing machine learning methodology to predict response to treatments based on brain scans alongside information such as age, stroke severity and pre-existing illnesses. Improving the accuracy of decision making is important for effective resource allocation and informing discussions between clinicians and families or carers. Expediting the decision-making process will also help reduce disparities in treatment provision across different areas and diverse patient groups.

Computers can be trained to identify patterns in large sets of information using machine learning (ML). Our project applies these tools on brain scans and clinical information extracted from electronic health records to improve prediction of how an individual patient will respond to treatment. We are currently developing this approach in a group of nearly 700 stroke patients who received thrombolysis or thrombectomy for stroke at Addenbrooke’s hospital.

A data-driven approach

Our project integrates brain scans and clinical data extracted from electronic health records linking with the Cardiovascular eHospital Research Database (eCamCVD).

A key aim of the project was to streamline data retrieval and establish workflows to allow further analysis and application of ML models. The Accelerate - C2D3 funding made this step possible by enabling us to hire a dedicated research associate to work on the project and establish the data pipeline central to our work. Furthermore, the imaging data was standardised and refined to enable application of machine learning models which we are currently exploring.

Our initial results indicate that these computational models can help extract predictive information from brain scans and improve accuracy of outcome prediction when combined with simple baseline clinical information.

While integrating a decision support tool into routine clinical practice comes with a number of logistic and regulatory challenges, these results are a good step towards making our goal possible in the future.

Looking ahead

There are a number of challenges that we need to be mindful of. The information we are utilising was collected as part of routine clinical care and not for a dedicated research trial. This makes the data noisy and this may introduce a degree of bias based on current practice. Our plan is to test our approach in two independent datasets to minimise risk of bias and increase reliability and generalisability of our results.

We will also test our approach in a different group of patients across East Anglia and more widely through an international collaboration with colleagues in Australia. The C2D3 funding has been instrumental in helping us set the groundwork for these wider collaborations.

An artificial intelligence-derived decision support tool to predict outcomes from early stroke treatments patients would help clinicians and could be used to give individual patients and their families a better idea of outcomes, so we can provide the best and targeted care from the start.

This project was funded though the 2023 Accelerate-C2D3 funding call for novel applications of AI for research and innovation. You can read more about other funded projects here.